This PDF file contains the front matter associated with SPIE Proceedings Volume 9443, including the Title Page, Copyright information, Table of Contents, and Conference Committee listing.

One of the key issues of face recognition is to extract the features of face images. In this paper, we propose a novel method, named two-dimensional discriminant neighborhood preserving embedding (2DDNPE), for image feature extraction and face recognition. 2DDNPE benefits from four techniques, i.e., neighborhood preserving embedding (NPE), locality preserving projection (LPP), image based projection and Fisher criterion. Firstly, NPE and LPP are two popular manifold learning techniques which can optimally preserve the local geometry structures of the original samples from different angles. Secondly, image based projection enables us to directly extract the optimal projection vectors from twodimensional image matrices rather than vectors, which avoids the small sample size problem as well as reserves useful structural information embedded in the original images. Finally, the Fisher criterion applied in 2DDNPE can boost face recognition rates by minimizing the within-class distance, while maximizing the between-class distance. To evaluate the performance of 2DDNPE, several experiments are conducted on the ORL and Yale face datasets. The results corroborate that 2DDNPE outperforms the existing 1D feature extraction methods, such as NPE, LPP, LDA and PCA across all experiments with respect to recognition rate and training time. 2DDNPE also delivers consistently promising results compared with other competing 2D methods such as 2DNPP, 2DLPP, 2DLDA and 2DPCA.

Face recognition based on machine vision has achieved great advances and been widely used in the various fields. However, there are some challenges on the face recognition, such as facial pose, variations in illumination, and facial expression. So, this paper gives the recent advances in 3D face recognition. 3D face recognition approaches are categorized into four groups: minutiae approach, space transform approach, geometric features approach, model approach. Several typical approaches are compared in detail, including feature extraction, recognition algorithm, and the performance of the algorithm. Finally, this paper summarized the challenge existing in 3D face recognition and the future trend. This paper aims to help the researches majoring on face recognition.

Supervised Descent Method (SDM) learns the descent directions of nonlinear least square objective in a supervised manner, which has been efficiently used for face alignment. However, SDM still may fail in the cases of partial occlusions and serious pose variations. To deal with this issue, we present a new method for robust face alignment by utilizing the low rank prior of human face and enforcing sparse structure of the descent directions. Our approach consists of low rank face frontalization and sparse descent steps. Firstly, in terms of the low rank prior of face image, we recover such a low-rank face from its deformed image and the associated deformation despite significant distortion and corruption. Alignment of the recovered frontal face image is more simple and effective. Then, we propose a sparsity regularized supervised descent model by enforcing the sparse structure of the descent directions under the l1constraint, which makes the model more effective in computation and robust to partial occlusion. Extensive results on several benchmarks demonstrate that the proposed method is robust to facial occlusions and pose variations

Face detection is an important aspect for biometrics, video surveillance and human computer interaction. Due to the complexity of the detection algorithms any face detection system requires a huge amount of computational and memory resources. In this communication an accelerated implementation of MB LBP face detection algorithm targeting low frequency, low memory and low power embedded system is presented. The resulted implementation is time deterministic and uses a customizable AMBA IP hardware accelerator. The IP implements the kernel operations of the MB-LBP algorithm and can be used as universal accelerator for MB LBP based applications. The IP employs 8 parallel MB-LBP feature evaluators cores, uses a deterministic bandwidth, has a low area profile and the power consumption is ~95 mW on a Virtex5 XC5VLX50T. The resulted implementation acceleration gain is between 5 to 8 times, while the hardware MB-LBP feature evaluation gain is between 69 and 139 times.

In this paper, we address the problem of three dimensional human pose tracking and estimation using Particle Swarm Optimization (PSO) with an improved silhouette extraction mechanism. In this work, the tracking problem is formulated as a nonlinear function optimization problem so the main objective is to optimize the fitness function between the 3D human model and the image observations. In order to improve the tracking performance, new shadow detection, removal and a level-set mechanism are applied during silhouette extraction. Both the silhouette and edge likelihood are used in the fitness function. Experiments using HumanEva-II dataset demonstrate that the proposed approach performance is considerably better than baseline algorithm which uses the Annealed Particle Filter (APF).

In this work, we propose two improvements of the Gestalt Interest Points (GIP) algorithm for the recognition of faces of people that have underwent significant weight change. The basic assumption is that some interest points contribute more to the description of such objects than others. We assume that we can eliminate certain interest points to make the whole method more efficient while retaining our classification results. To find out which gestalt interest points can be eliminated, we did experiments concerning contrast and orientation of face features. Furthermore, we investigated the robustness of GIP against image rotation. The experiments show that our method is rotational invariant and - in this practically relevant forensic domain - outperforms the state-of-the-art methods such as SIFT, SURF, ORB and FREAK.

This paper introduces a method for facial expression recognition combining appearance and geometric facial features. The proposed framework consistently combines multiple facial representations at both global and local levels. First, covariance descriptors are computed to represent regional features combining various feature information with a low dimensionality. Then geometric features are detected to provide a general facial movement description of the facial expression. These appearance and geometric features are combined to form a vector representation of the facial expression. The proposed method is tested on the CK+ database and shows encouraging performance.

This paper addresses the problem of retail product recognition on grocery shelf images. We present a technique for accomplishing this task with a low time complexity. We decompose the problem into detection and recognition. The former is achieved by a generic product detection module which is trained on a specific class of products (e.g. tobacco packages). Cascade object detection framework of Viola and Jones [1] is used for this purpose. We further make use of Support Vector Machines (SVMs) to recognize the brand inside each detected region. We extract both shape and color information; and apply feature-level fusion from two separate descriptors computed with the bag of words approach. Furthermore, we introduce a dataset (available on request) that we have collected for similar research purposes. Results are presented on this dataset of more than 5,000 images consisting of 10 tobacco brands. We show that satisfactory detection and classification can be achieved on devices with cheap computational power. Potential applications of the proposed approach include planogram compliance control, inventory management and assisting visually impaired people during shopping.

This Research study proposes a novel method for face recognition based on Anthropometric features that make use of an integrated approach comprising of a global and personalized models. The system is aimed to at situations where lighting, illumination, and pose variations cause problems in face recognition. A Personalized model covers the individual aging patterns while a Global model captures general aging patterns in the database. We introduced a de-aging factor that de-ages each individual in the database test and training sets. We used the k nearest neighbor approach for building a personalized model and global model. Regression analysis was applied to build the models. During the test phase, we resort to voting on different features. We used FG-Net database for checking the results of our technique and achieved 65 percent Rank 1 identification rate.

In this paper, we investigated the efficiency of upper half Farsi numerical digit structure. In other words, half of data (upper half of the digit shapes) was exploited for the recognition of Farsi numerical digits. This method can be used for both offline and online recognition. Half of data is more effective in speed process, data transfer and in this application accuracy. Hidden Markov model (HMM) was used to classify online Farsi digits. Evaluation was performed by TMU dataset. This dataset contains more than 1200 samples of online handwritten Farsi digits. The proposed method yielded more accuracy in recognition rate.

A novel and simple ellipse detection method is proposed in this paper. First, Canny operator is carried on the gray image to get edge image. Second, all the edge segments are obtained from edge image and output gradients of edge segments for further analysis. According to gradient direction, the edge segments are split into primitive lines and arcs. Then elliptical arcs are extracted from the results of splitting and an efficient grouping strategy is proposed to group elliptical arcs coming from the same ellipse as candidate ellipse. Finally, least-square fitting method is implemented to estimate the parameters of these candidate ellipses. Experiment results show that the proposed method is robust to noise and fast for real-time implementation.

In this paper, we propose a stereo vision-based pedestrian detection method using a dense disparity map-based detection and segmentation algorithm. To enhance a pedestrian detection performance, we use a dense disparity map extracted from a global stereo matching algorithm. First, we extract a road feature information from the dense disparity map, which is a decision basis of presence or absence of obstacles on the road. It is very important to extract the road feature from the disparity for detecting obstacles robustly regardless of external traffic situations. The obstacle detection is performed with the road feature information to detect only obstacles from entire image. In other words, pedestrian candidates including various upright objects are detected in the obstacle detection stage. Each obstacle area tends to include multiple objects. Thus, a disparity map-based segmentation is performed to separate the obstacle area into each obstacle accurately. And then, accurate pedestrian areas are extracted from segmented obstacle areas using road contact and pedestrian height information. This stage enables to reduce false alarms and to enhance computing speed. To recognize pedestrians, classifier is performed in each verified pedestrian candidate. Finally, we perform a verification stage to examine the recognized pedestrian in detail. Our algorithms are verified by conducting experiments using ETH database.

Lane detection plays a significant role in Advanced Driver Assistance Systems (ADAS) for intelligent vehicles. In this paper we present a multi-lane detection method based on multiple vanishing points detection. A new multi-lane model assumes that a single lane, which has two approximately parallel boundaries, may not parallel to others on road plane. Non-parallel lanes associate with different vanishing points. A biological plausibility model is used to detect multiple vanishing points and fit lane model. Experimental results show that the proposed method can detect both parallel lanes and non-parallel lanes.

Optical Mark Recognition (OMR) is a traditional data input technique and an important human computer interaction technique which is widely used in education testing. This paper proposes a new idea for grading multiple-choice test which is based on a camera on smartphone. The system key techniques and relevant implementations, which include the image scan, edge detection and reduction of brightness on colorful bubble form images, are presented.

The waste of fish food has always been a serious problem in aquaculture. On one hand, the leftover fish food spawns a big waste in the aquaculture industry because fish food accounts for a large proportion of the investment. On the other hand, the left over fish food may pollute the water and make fishes sick. In general, the reason for fish food waste is that there is no feedback about the consumption of delivered fish food after feeding. So it is extremely difficult for fish farmers to determine the amount of feedstuff that should be delivered each time and the feeding intervals. In this paper, we propose an effective method using image processing techniques to solve this problem. During feeding events, we use an underwater camera with supplementary LED lights to obtain images of uneaten fish food pellets on the tank bottom. An algorithm is then developed to figure out the number of left pellets using adaptive Otsu thresholding and a linear-time component labeling algorithm. This proposed algorithm proves to be effective in handling the non-uniform lighting and very accurate number of pellets are counted in experiments.

In order to meet the real-time requirement of video copy detection, a robust video copy detection algorithm is proposed. Harris feature points is extracted based on local feature descriptor, video frames are divided into blocks and video fingerprints are generated by calculating video feature point amplitude and angle differences. The matching result graph is formed based on video matching frames, copy video is detection by searching the longest path of the graph. Compared with other video detection algorithm, proposed algorithm is with good robustness and discrimination accuracy, experiment proved that detection speed is improved further.

Robust detection of moving objects in image sequences is an essential part of many vision applications. However, it is not easily achievable with a moving camera since the camera and moving objects motions are mixed together. In this paper we propose a method to detect moving objects under a moving camera. The camera ego-motion is compensated by the corresponding feature sets. The difference image between two consecutive images that ego-motion is compensated is transformed into a binary image using k-means algorithm. According to the clustering results, the region of interest where moving objects are likely to exist is searched by the projection approach. Then local threshold and contour filling methods are applied to detect the accurate moving objects. Experimental results on real image sequences demonstrate that our method can get intact moving objects in the case of a moving camera efficiently.

Human abnormal behaviors detection is one of the most challenging tasks in the video surveillance for the public security control. Interaction Energy Potential model is an effective and competitive method published recently to detect abnormal behaviors, but their model of abnormal behaviors is not accurate enough, so it has some limitations. In order to solve this problem, we propose a novel Particle Motion model. Firstly, we extract the foreground to improve the accuracy of interest points detection since the complex background usually degrade the effectiveness of interest points detection largely. Secondly, we detect the interest points using the graphics features. Here, the movement of each human target can be represented by the movements of detected interest points of the target. Then, we track these interest points in videos to record their positions and velocities. In this way, the velocity angles, position angles and distance between each two points can be calculated. Finally, we proposed a Particle Motion model to calculate the eigenvalue of each frame. An adaptive threshold method is proposed to detect abnormal behaviors. Experimental results on the BEHAVE dataset and online videos show that our method could detect fight and robbery events effectively and has a promising performance.

This article proposes a novel hybrid motion detection algorithm based on 2-D (2-Dimensional) spatio-temporal states histogram. The new algorithm combines the idea of image change detection based on 2-D histogram and spatio-temporal entropy image segmentation. It quantifies the continuity of pixel state in time and space domain which are called TDF (Time Domain Filter) and SDF (Space Domain Filter) respectively. After this, put both channels of output data from TDF and SDF into a 2-D histogram. In the 2-D histogram, a curve division method helps to separate the foreground state points and the background ones more accurately. Innovatively, the new algorithm converts the video sequence to its histogram sequence, and transforms the difference of pixel’s value in the video sequence into the difference of pixel’s position in the 2-D histogram. Experimental results on different types of scenes added Gaussian noise shows that the proposed technique has strong ability of detecting moving objects.

Detecting dim and small target in infrared images and videos is one of the most important techniques in many computer vision applications, such as video surveillance and infrared imaging precise guidance. In this paper, we proposed a real-time target detection approach in infrared imagery. This method combined saliency detection technology and local average filtering. First, we compute the log amplitude spectrum of infrared image. Second, we find the spikes of the amplitude spectrum using cubic facet model and suppress the sharp spikes using local average filtering. At last, the detection result in spatial domain is obtained by reconstructing the 2D signal using the original phase and the filtered amplitude spectrum. Experimental results of infrared images with different types of backgrounds demonstrate the high efficiency and accuracy of the proposed method to detect the dim and small targets.

In this paper, we propose an alternative salient object detection method based on maximum saliency region and density. The proposed approach can automatically detect the salient object with a well-defined boundary. Saliency region and density maximization is used as the quality function to find the optimal window containing a salient object. And for efficiently executing window search, a branch-and-bound search algorithm based on saliency region and density is proposed. Then the located window is used to initialize the GrabCut method, and the salient object with a well- defined boundary is extracted through applying GrabCut. Experimental results show that the proposed salient object detection approach outperforms the state-of-the-art methods.

This paper proposes a framework for small infrared target real-time visual enhancement. The framework is consisted of three parts: energy accumulation for small infrared target enhancement, noise suppression and weighted fusion. Dynamic programming based track-before-detection algorithm is adopted in the energy accumulation to detect the target accurately and enhance the target’s intensity notably. In the noise suppression, the target region is weighted by a Gaussian mask according to the target’s Gaussian shape. In order to fuse the processed target region and unprocessed background smoothly, the intensity in the target region is treated as weight in the fusion. Experiments on real small infrared target images indicate that the framework proposed in this paper can enhances the small infrared target markedly and improves the image’s visual quality notably. The proposed framework outperforms tradition algorithms in enhancing the small infrared target, especially for image in which the target is hardly visible.

Breast cancer is one of the types of cancer which is most commonly seen in women. Density of breast is an important indicator for the risk of cancer. In addition, densities of tissue may harden the diagnosis by hiding the abnormalities occurring on the breast. For this reason, during the process of diagnosis, the process of automatic classification of breast density has a significant importance. In this study, a new system with the base of Artificial Neural Network (ANN) and multiple resolution analysis is suggested. Wavelet and curvelet analyses having the most common use have been used as multi resolution analysis. 4 pieces of statistics which are minimum value, maximum value, mean value and standard deviation have been extracted from the images which have been eluted to their sub-bands via multi resolution analysis. For the purpose of testing the success of the system, 322 pieces of images which are in MIAS database have been used. The obtained results for different backgrounds are so satisfying; and the highest classification values have been obtained as 97.16 % with Wavelet transform and ANN for fatty background and 79.80 % with Wavelet transform and ANN for fatty-glanduar background. The same results have been obtained using Wavelet transform and ANN and Curvelet transform and ANN for dense background and accuracy rate of 84.82 % have been reached. The results of mean classification have been obtained, for three pieces of tissue types (fatty, fatty-glanduar, dense), in sequence as 84.47 % with the use of ANN, 85.71 % with the use of curvelet analysis and ANN; and 87.26 % with the use of wavelet analysis and ANN.

We proposed an adaptive method for more efficient distance-coding of binary images. The proposed method partitions the image into blocks where the interval sequences of zeros or ones can be calculated, as opposed to the conventional method where intervals are calculated by following a fixed scan order. In the proposed method, one can adaptively choose either horizontal or vertical scan within a block, depending on criteria based on entropy values. The resulting intervals tend to have lower entropies than the conventional non-adaptive methods, thereby allowing for higher compression than distance coding using a lossless codec. Our simulations on various test images demonstrated that (i) the proposed method achieved significantly higher compression than the non-adaptive distance coding method; (ii) the proposed method can be used as an efficient preprocessor of a lossless coder, offering higher compression than directly coding on the original images.

To recognize, track and count the bolting machine in coal mine video images, a real-time target tracking method based on the Lucas-Kanade sparse optical flow is proposed in this paper. In the method, we judge whether the moving target deviate from its trajectory, predicate and correct the position of the moving target. The method solves the problem of failure to track the target or lose the target because of the weak light, uneven illumination and blocking. Using the VC++ platform and Opencv lib we complete the recognition and tracking. The validity of the method is verified by the result of the experiment.

To improve tracking drift which often occurs in adaptive tracking, an algorithm based on the fusion of tracking and detection is proposed in this paper. Firstly, object tracking is posed as abinary classification problem and is modeled by partial least squares (PLS) analysis. Secondly, tracking object frame by frame via particle filtering. Thirdly, validating the tracking reliability based on both positive and negative models matching. Finally, relocating the object based on SIFT features matching and voting when drift occurs. Object appearance model is updated at the same time. The algorithm can not only sense tracking drift but also relocate the object whenever needed. Experimental results demonstrate that this algorithm outperforms state-of-the-art algorithms on many challenging sequences.

The precious battle damage assessment by use of video images to analysis missile placement is a new study area. The article proposed an improved speeded up robust features algorithm named restricted speeded up robust features, which combined the combat application of TV-command-guided missiles and the characteristics of video image. Its restrictions mainly reflected in two aspects, one is to restrict extraction area of feature point; the second is to restrict the number of feature points. The process of missile placement analysis based on video image was designed and a video splicing process and random sample consensus purification were achieved. The RSURF algorithm is proved that has good realtime performance on the basis of guarantee the accuracy.

Temperature is an important feature of infrared targets. However, due to the attenuation and distortion parameters in radiation transmission process are unknown, precise temperature measurement is a difficult task. In this paper, a modified Dual-Band Ratio (DBR) temperature measurement method for remote target is proposed. The method is based on a new presented variation derived from the temperature change process named Dual-Band Differential Ratio (DBDR). Firstly, the temperature of the target is estimated by the traditional DBR method, and then a correction using DBDR information is carried out to improve the measurement accuracy. Experiment results showed that the proposed method can improve the temperature measurement accuracy and it could also be carried out without any prior information about the target.

The current performance requirements in soccer make imperative the use of new technologies for game observation and analysis, such that detailed information about the teams’ actions is provided. This paper summarizes a framework to collect the soccer players and ball positions using one or more Full HD Handycams, placed no more than 20cm apart in the stands, as well as how this framework connects to the FootData project. The system was based on four main modules: the detection and delimitation of the soccer pitch, the ball and the players detection and assignment to their teams, the tracking of players and ball and finally the computation of their localization (in meters) in the pitch.

In this paper we present a combination of block-matching and differential motion field estimation. We initialize the motion field using a predictive hierarchical block-matching approach. This vector field is refined by a pixel-recursive differential motion estimation method. We integrate image warping and adaptive filter kernels into the Horn and Schunck differential optical flow estimation approach to break the block structure of the initial correspondence vector fields and compute motion field updates to fulfill the smoothness constraint inside motion boundaries. The influence of occlusion areas is reduced by integrating an in-the-loop occlusion detection and adjusting the adaptive filter weights in the iteration process. We integrate the combined estimation into a hierarchical multi-scale framework. The refined motion on the current scale is upscaled and used as prediction for block-matching motion estimation on the next scale. With the proposed system we are able to combine the advantages of block-matching and differential motion estimation and achieve a dense vector field with floating point precision even for large motion.

In this paper, we describe a novel method for visual vehicle tracking process based on the combination of speeded-up robust features (SURF) points and color feature. The whole tracking process is constructed in the framework of particle filter. To further improve the precision and stability of tracking, a dynamic update mechanism of target template is proposed to capture appearance changes. This mechanism includes two strategies: Adopting new feature points and discarding bad feature points. A novel distance kernel function method is adopted to allocate the weight of each particle, and to improve the stability of the tracking template. The experiments present that our algorithm can track the targets more robustly and adaptively than the traditional algorithms.

Protecting fine details and edges while denoising digital images is a challenging area of research due to changing characteristics of both, noise and signal. Denoising is used to remove noise from corrupted images but in the process fine details like weak edges and textures are hampered. In this paper we propose an algorithm based on Ridgelet transform to denoise images and protect fine details. Here we use cycle spinning on Ridgelet coefficients with soft thresholding and name the algorithm as Ridgelet Shrinkage in order to suppress noise and preserve details. The projections in Ridgelets filter out the noise while protecting the details while the ridgelet shrinkage further suppress noise. The proposed algorithm out performs the Wavelet Shrinkage and Non-local (NL) means denoising algorithms on the basis of Peak Signal to Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM) numerically and visually both.

In this paper, we propose a novel image super-resolution method based on sparse principle component analysis. Various coupled sub-dictionaries are trained to represent high-resolution and low-resolution image patches. The proposed method simultaneously exploits the incoherence of the sub-dictionaries and nonlocal self-similarity existing in natural images. The purpose of introducing these two regularization terms is to design a novel dictionary learning algorithm for having good reconstruction. Furthermore, in the dictionary learning process, the algorithm can update the dictionary as a whole and reduce the computational cost significantly. Experimental results show the efficiency of the proposed method compared to the existing algorithms in terms of both PSNR and visual perception.

The main aim of the paper is to present a non-linear image quality assessment model based on a fuzzy logic estimator, namely the Takagi-Sugeno-Kang fuzzy model. This image quality assessment model uses a clustered space of input objective metrics. Main advantages of the introduced quality model are simplicity and understandably of its fuzzy rules. As reference model the polynomial 3 rd order model was chosen. The parameters of the Takagi-Sugeno-Kang fuzzy model are optimized in accordance to the mapping criteria of the selected set of input objective quality measures to the Mean Opinion Score (MOS) scale.

Markov random field (MRF) has been widely used in SAR image segmentation because of the advantage of directly modeling the posterior distribution and suppresses the speckle on the influence of the segmentation result. However, when the real SAR images are nonstationary images, the unsupervised segmentation results by MRF can be poor. The recent proposed triplet Markov field (TMF) model is well appropriate for nonstationary SAR image processing due to the introduction of an auxiliary field which reflects the nonstationarity. In addition, on account of the texture features of SAR image, a fusion image segmentation method is proposed by fusing the gray level image and texture feature image. The effectiveness of the proposed method in this paper is demonstrated by a synthesis SAR image and the real SAR images segmentation experiments, and it is better than the state-of-art methods.

In this paper, we present a method for segmentation of document page flow applied to heterogeneous real bank documents. The approach is based on the content of images and it also incorporates font based features inside the documents. Our method involves a bag of visual words (BoVW) model on the designed image based feature descriptors and a novel approach to combine the consecutive pages of a document into a single feature vector that represents the transition between these pages. The transitions here could be represented by one of the two different classes: continuity of the same document or beginning of a new document. Using the transition feature vectors, we utilize three different binary classifiers to make predictions on the relationship between consecutive pages. Our initial results demonstrate that the proposed method can exhibit promising performance for document flow segmentation at this stage.

In this paper a novel method for blind noisy image quality assessment is proposed. First, it is believed that human visual system (HVS) is more sensitive to the local smoothness area in a noise image, an adaptively local homogeneous block selection algorithm is proposed to construct a new homogeneous image named as homogeneity blocks (HB) based on computing each pixel characteristic. Second, applying the discrete cosine transform (DCT) for each HB and using high frequency component to evaluate image noise level. Finally, a modified peak signal to noise ratio (MPSNR) image quality assessment approach is proposed based on analysis DCT kurtosis distributions change and noise level above-mentioned. Simulations show that the quality scores that produced from the proposed algorithm are well correlated with the human perception of quality and also have a stability performance.

Image segmentation is critical for image processing. Among several algorithms, GrabCut is well known by its little user interaction and desirable segmentation result. However, it needs to take a lot of time to adjust the Gaussian Mixture Model (GMM) and to cut the weighted graph with Max-Flow/Min-Cut Algorithm iteratively. To solve this problem, we first build a common algorithmic framework which can be shared by the class of GrabCut-like segmentation algorithms, and then propose KM_GrabCut algorithm based on this framework. The KM_GrabCut first uses K-means clustering algorithm to cluster pixels in foreground and background respectively, and then constructs a GMM based on each clustering result and cuts the corresponding weighted graph only once. Experimental results demonstrate that KM_GrabCut outperforms GrabCut with higher performance, comparable segmentation result and user interaction.

It has been realized an estimation of variance of the sea surface slopes through the variances on images that consist of bright and dark regions that are called glitter pattern. The probability distribution of the sea surface slopes has been used considering a non-Gaussian case taking in account the skewness and the kurtosis of the sea surface slopes. These relationships of variance have been calculated for five different angles of light incidence on the sea surface and for four different heights of the image sensor. The brightness in the glittern pattern has been modeled using a Gaussian function with information of the incident and reflection light angle in its argument. Some computational aspects and applications for optical engineering are mentioned.

A novel image segmentation algorithm which uses quantum entropy and pulse-coupled neural networks (PCNN) is proposed in this paper. Optimal iteration of the PCNN is one of the key factors affecting segmentation accuracy. We borrow quantum entropy from quantum information to act as a criterion in determining optimal iteration of the PCNN. Optimal iteration is captured while total quantum entropy of the segments reaches a maximum. Moreover, compared with other PCNN-employed algorithms, the proposed algorithm works without any manual intervention, because all parameters of the PCNN are set automatically. Experimental results prove that the proposed method can achieve much lower probabilities of error segmentation than other PCNN-based image segmentation algorithms, and this suggests that higher image segmentation quality is achieved by the proposed method.

New algorithms are proposed in this paper for single image super-resolution using multiple dictionaries based on sparse representation. In the proposed algorithms, a classifier is constructed which is based on the edge properties of image patches via the two lowest discrete cosine transformation (DCT) coefficients. The classifier partitions all training patches into three classes. Training patches from each of the three classes can then be used for the training of the corresponding dictionary via the K-SVD (singular value decomposition) algorithm. Experimental results show that the high resolution image quality using the proposed algorithms is better than that using the traditional bi-cubic interpolation and Yang’s method.

It is proposed of an improved median de-noising method, namely an image de-noising algorithm based on clustering and median filtering. The algorithm is a kind of image fast de-noising method based on the clustering idea, the singular point points are isolated from the image and then clustering. It is advantage to better protect the details of an image and to substantially reduce calculation. Compared with traditional median filter, mean filter and wiener filter, our approach is more adaptive and receives better results. While for images that have complex details such as texture images, the results of experiment show that the proposed algorithm works less well in the de-noising effect comparatively.

Block matching 3-D transform (BM3D) is an excellent image denoising algorithm, the success of this algorithm is that it makes full use of the image self-similarity. Profiting from the superior performance of the BM3D, this paper propose a multi-focus image fusion algorithm. A pair of the original images are both implemented BM3D respectively, averaging their low frequency coefficients can get the low frequency coefficient of the result image, the bigger high frequency coefficient in these two transforms is chosen as the high frequency coefficient of the result image; Implementing inverse 3-D transformation can obtain the fused image. Experimental results show that the proposed algorithm is better than the existing multi-focus image fusion algorithms on both the subjective visual and the objective evaluation.

In order to increase the hiding capacity and provide visually imperceptible changes of an image, a multi-way pixelvalue differencing scheme is proposed in this paper. A threshold value is designated to distinguish between edge-like and smooth blocks. The smooth blocks adopt the original tri-way pixel-value differencing (TPVD) method whereas the edgelike blocks apply the proposed mode selection algorithm to further determine the data embedding way by utilizing horizontal or vertical edges. In addition, a further altering process is conducted to achieve the original partitioning results after the embedding procedure so that the stego-images can extract secret data without the participation of the original cover images. The experimental results show that the proposed method can embed a large hiding capacity in a cover image and the image quality remains high.

In this paper, we propose a region of interest-based (ROI-adaptive) fusion algorithm of infrared and visible images by using the Laplacian Pyramid method. Firstly, we estimate the saliency map of infrared images, and then divide the infrared image into two parts: the regions of interest (RoI) and the regions of non-interest (nRoI), by normalizing the saliency map. Visible images are also segmented into two parts by using the Gauss High-pass filter: the regions of high frequency (RoH) and the regions of low frequency (RoL). Secondly, we down-sampled both the nRoI of infrared image and the RoL of visible image as the input of next level processing. Finally, we use normalized saliency map of infrared images as the weighted coefficient to get the basic image on the top level and choose max gray value of the RoI of infrared image and the RoH of visible image to get the detail image. In this way, our method can keep target feature of infrared image and texture detail information of visual image at the same time. Experiment results show that such fusion scheme performs better than the other fusion algorithms both on human visual system and quantitative metrics.

Applying fractional Fourier transform (FRFT) on an image, the phase and amplitude portion have different capability to reflect spectral information of the source image. Generally, the phase portion is of more importance. Taking advantage of these characters, a novel image fusion algorithm, named as FRFT-phase-amplitude, is proposed. Firstly, apply FRFT on source images, and then the amplitude and the phase information are separated. Secondly, fuse the amplitude portion in FRFT domain using the largest absolute value fusion rule. Thirdly, do inverse fractional Fourier transform (IFRFT) on phase portions to get reconstructed phase images, and fuse them in spatial domain by selecting the larger pixel value, then process FRFT on this fused phase image. Finally, combine fused phase portion with fused amplitude portion in fractional domain, and apply IFRFT on the combination to create the fused image. Experiments reveal that the FRFT-phase- amplitude algorithm can produce better fusion effects than methods based on wavelet transform and FRFT.

During model-based image dehazing, the role of the accuracy of transmission estimation is crucial, which has a decisive effect on the final result. Considering that an ideal transmission map must be smooth, edge-preserving and free of redundant false details, a fusion-based dark channel prior (DCP) dehazing algorithm is presented in this paper. On the basis of DCP, a pixel-wise and a patch-wise transmission maps are obtained. Then an L0 smoothing filter and a large scale Gaussian filter are applied to them respectively. Finally, a much more accurate refined transmission map is attained through fusion and a haze-free image is restored using the atmosphere degradation model. Furthermore, a novel scheme for setting the lower bound of transmission adaptively is also put forward. Experiments demonstrate a better and faster dehazing capability over original DCP algorithm and state-of-the-art dehazing methods, especially in suppressing halo artifacts, restoring details and coping with the haze existing in small-scale areas of depth discontinuity occluded by foreground..

A contrast correction method is presented based on nonlinear mapping of windowed tone. The main idea of method is to employ the local nonlinear mapping model on the small size with overlapping windows of traversal the whole image. At first, a high dynamic range (HDR) image contrast correction is introduced, and then through the formula deduction, a model for decision optimization of contrast correction is established, in which some constraints are termed as two adaptive guided images based on human visual properties so as to improve the optimal solution. Finally, the optimal contrast correction can be implemented by solving the optimizing processing problem through a linearized reduction. A series of experiments with the HDR natural images are carried out and the results of objective quality metrics have showed that the proposed method can effectively improve and optimize the contrast correction to outperform those current existing methods.

Histogram specification is a useful algorithm of image enhancement field. This paper proposes an image haze removal algorithm of histogram specification based on the weighted Gaussian probability density function (Gaussian PDF). Firstly, we consider the characteristics of image histogram that captured when sunny, fogging and haze weather. Then, we solve the weak intensity of image specification through changing the variance and weighted Gaussian PDF. The performance of the algorithm could removal the effective of fog and experimental results show the superiority of the proposed algorithm compared with histogram specification. It also has much advantage in respect of low computational complexity, high efficiency, no manual intervention.

Image matting is an important task in image and video editing and has been studied for more than 30 years. In this paper we propose an improved interactive matting method. Starting from a coarse user-guided trimap, we first perform a color estimation based on texture and color information and use the result to refine the original trimap. Then with the new trimap, we apply soft matting process which is improved Bayesian matting with smoothness constraints. Experimental results on natural image show that this method is useful, especially for the images have similar texture feature in the background or the images which is hard to give a precise trimap.

This paper presents a genetic algorithm for bundle adjustment in aerial panoramic stitching. Compared with the conventional LM (Levenberg-Marquardt) algorithm for bundle adjustment, the proposed bundle adjustment combining the genetic algorithm optimization eliminates the possibility of sticking into the local minimum, and not requires the initial estimation of desired parameters, naturally avoiding the associated steps, that includes the normalization of matches, the computation of homography transformation, the calculations of rotation transformation and the focal length. Since the proposed bundle adjustment is composed of the directional vectors of matches, taking the advantages of genetic algorithm (GA), the Jacobian matrix and the normalization of residual error are not involved in the searching process. The experiment verifies that the proposed bundle adjustment based on the genetic algorithm can yield the global solution even in the unstable aerial imaging condition.

In this paper, a novel fragile image watermarking scheme based on singular value decomposition (SVD) using genetic algorithm (GA) is proposed. Every line of watermark is scaled by using multiple scaling factors (SFs). Host image is divided into blocks. Watermarked image is obtained by embedding a different line of the watermark to singular values (SVs) of the every block. In this proposed method, the SFs are optimized using GA to obtain maximum transparency. Experimental results indicate that the method reached the highest possible transparency. Fragility of the watermark under various attacks such as rotating, rescaling and sharpening is tested. When an attack does not occur, exactly the original extracted watermark is obtained; on the other hand, the extracted watermark is intensely distorted.

Content-based image retrieval (CBIR) has become an interesting and urgent research topic due to the increase of necessity of indexing and classification of multimedia content in large databases. The low level visual descriptors, such as color-based, texture-based and shape-based descriptors, have been used for the CBIR task. In this paper we propose a color-based descriptor which describes well image contents, integrating both global feature provided by dominant color and local features provided by color correlogram. The performance of the proposed descriptor, called Dominant Color Correlogram descriptor (DCCD), is evaluated comparing with some MPEG-7 visual descriptors and other color-based descriptors reported in the literature, using two image datasets with different size and contents. The performance of the proposed descriptor is assessed using three different metrics commonly used in image retrieval task, which are ARP (Average Retrieval Precision), ARR (Average Retrieval Rate) and ANMRR (Average Normalized Modified Retrieval Rank). Also precision-recall curves are provided to show a better performance of the proposed descriptor compared with other color-based descriptors.

Glasses removal is an important task on face recognition, in this paper, we provide a weakly supervised method to remove eyeglasses from an input face image automatically. We choose sparse coding as face reconstruction method, and optical flow to find exact shape of glasses. We combine the two processes iteratively to remove glasses more accurately. The experimental results reveal that our method works much better than these algorithms alone, and it can remove various glasses to obtain natural looking glassless facial images.

In order to enhance the local feature’s describing capacity and improve the classification performance of high-resolution (HR) satellite images, we present an HR satellite image scene classification method that make use of spatial information of local feature. First, the spatial pyramid matching model (SPMM) is adopted to encode spatial information of local feature. Then, images are represented by the local feature descriptors and encoding information. Finally, the support vector machine (SVM) classifier is employed to classify image scenes. The experiment results on a real satellite image dataset show that our method can classify the scene classes with an 82.6% accuracy, which indicates that the method can work well on describing HR satellite images and classifying different scenes.