This is a list of reviewers who served the Journal of Electronic Imaging in 2019.

We propose a parallel network with spatial–temporal attention for video-based person re-identification. Many previous video-based person re-identification methods use two-dimensional convolutional neural networks to extract spatial features, then, temporal features are extracted by temporal pooling or recurrent neural networks. Unfortunately, these series networks will cause the loss of spatial information when extracting temporal information. Different from previous methods, our parallel network can extract temporal and spatial features simultaneously, which can effectively reduce the loss of space information. In addition, we design a global temporal attention module, which obtains the attention weight through the correlation between the current frame and all the frames in the sequence. At the same time, the temporal module can act on the information extraction of spatial module. In this way, we can increase the temporal and spatial constraints. Experiments show that our method can effectively improve the re-id accuracy, better than the state-of-the-art methods.

Rate control plays a key role in high-efficiency video coding. For the λ domain rate control, the λ is calculated according to the allocated bits per pixel and by adaptively updating parameters. However, the selected λ is unsuitable for minimizing the distortion of the whole sequence. Considering the distortion effect in the temporal domain, a distortion propagation-based optimal λ decision algorithm is proposed. First, according to the reference ratio in the random access (RA) configuration, the distortion effect estimation function is derived with the RA main reference model. Second, based on the RDO processes of the sequence level, the group of picture level, and the frame level rate control, the special relationships of the current level λ and the down level λ are developed. Finally, with the results of the multilevel RDO processes, the temporal λ proportions are generated to obtain the reasonable λs for the different frames and coding tree units. The experimental results show that the coding performance of the proposed algorithm is significantly improved, with a good rate control accuracy.

Time–frequency analysis is a fundamental tool in many applications, such as inverse synthetic aperture radar (ISAR) imaging along the cross-range direction. Traditional time–frequency transformations designed for the general signal suffer low time–frequency resolution or cross-term interference. In this study, a cascaded UNet-like network is applied to the refinement of basic transformations, and another forward regression network is proposed to estimate the signal parameters directly. Both networks can incorporate a priori information and combine different time–frequency transformations efficiently. Several experiments, especially in the ISAR application, are presented to validate the methods. Through this research, the neural network is a promising approach to develop a customized method with high performance for a specific signal processing problem.

To achieve high yield of crops and to avoid pesticide abuse, different control methods have been adopted according to different degrees of disease in crop plants. To address the issue, a leaf disease grade identification method based on a convolutional neural network (CNN) was proposed. First, nonuniform illumination images were processed using an adaptive adjustment algorithm based on a two-dimensional (2-D) gamma function. Then a threshold segmentation method was used to segment diseased leaf images and thus to obtain binary images. Next, the ratio of the number of pixels in the lesion area to that in the diseased leaf area was calculated. This ratio was regarded as the classification threshold of the disease grades, and therefore was used to determine the disease grade category. In addition, use of a ResNet50-based CNN was proposed to identify disease grades, with a focal loss function replacing the standard cross entropy loss function, and with the Adam optimization method. Finally, leaf disease grade identification was performed on a database containing 10 types of disease leaf images for 8 crops, and it yielded a recognition accuracy of 95.61%. The experimental results showed that the proposed method was feasible and effective for the classification of leaf disease grades.

The judicious use of Compton scattered radiation continues to be a relevant challenge for the future of tomographic imaging. In this context, a new modality of Compton scattering tomography has been recently proposed. This new system has the advantage of being fixed, with a source and detectors placed on a ring passing through the source. The objective of this paper is to explore new possibilities for this new modality: scanning both small and large objects via a double scanning configuration. In fact, when an object under study is small and can be placed inside the detector ring, its scanning will be performed by the interior configuration. Otherwise, a large object placed outside the system will be scanned by the exterior configuration. The modeling of data acquisition of these two scanning modes leads to a new Radon transform on circular arcs (CirArcRT), established for the first time. Image reconstruction requires the inversion of CirArcRT, which is obtained via geometric inversion (GI). In fact, GI allows converting the circular arcs into corresponding half-lines, and consequently, the CirArcRT into a new Radon transform on half-lines, whose inversion is established. Numerical simulations are carried out to demonstrate the feasibility of the new modality. The ability for the same system to scan both small and large objects enlarges the fields of applications for tomography from biomedical imaging to nondestructive testing in industry as well as cultural heritage objects imaging.

The performance of face detection has achieved great progress with the rapid development of the convolutional neural network. However, detecting faces with a wide range of scales is still a challenging and crucial problem, especially for small faces. We focus on designing a strong network architecture, aiming at generating high-quality feature representation to handle the multiscale problem. Specifically, we conduct an in-depth study of three aspects, including feature fusion, attention mechanism, and receptive field, then propose an effective one-stage face detector. First, we design a feature fusion module, which adopts a more effective manner to sufficiently fuse multiscale features for building discriminative feature pyramid representation. Then, we propose an attention enhancement module for introducing the attention mechanism, which makes the network keep more valid information to emphasize the features from facial regions. Finally, we construct a prediction module, which ensures that each detection layer contains rich receptive fields to bring more useful context information. A large number of experiments on the widely used face detection benchmarks (i.e., WIDER FACE, FDDB, PASCAL FACE, and AFW) indicate that the results of our method outperform overwhelming majority of the existing methods and are comparable to the most recent state-of-the-art face detector, which demonstrate that our method is able to detect faces with a wide range of scales effectively.

In order to make full use of the spatial–spectral information of hyperspectral images (HSIs), a multiscale three-dimensional (3-D) convolutional neural network (CNN) model is proposed to extract the features of HSIs and classify them. First, the principal component analysis method is used to reduce the dimension of the original HSI, reduce the spectrum size on the basis of retaining the majority of spectrum information, and improve the training speed of the model. Second, at present, most of the HSI classification methods are based on spectral features and ignore the pixels’ neighborhood information on each band, thus a 3-D-CNN model is proposed for convolution feature extraction using a spatial–spectral joint method, which makes full use of the spatial–spectral information of HSI and greatly improves the classification effect of the model. Finally, traditional CNN models mostly use single-size convolution kernels. Several convolution kernels of different sizes are added to the convolution layer to extract the spatial features of different sizes, which effectively improves the classification ability of the model. The experimental results show that the multiscale 3-D-CNN model designed, when used on three datasets: Indian Pines, Pavia University, and Salinas, reached 97.54%, 99.78%, and 99.24% overall classification accuracy, respectively. Compared with the traditional single-scale two-dimensional-CNN model, the overall classification accuracy is improved by about 1% to 2%, which can better complete the classification task of HSIs.

Evaluating the performance of image fusion algorithms objectively is still an open problem. We proposed an integrational approach to objective image fusion evaluation based on integrating different existing objective metrics in order to overcome their individual shortcomings. The process of the objective measure construction contains two steps: first, we construct a candidate measure set, where each element contains high evaluation accuracy and low correlation with each other; then, the weights of measures in constructing the target measure are determined dynamically on training image set. Considering that the true quality of fused images is a subjective perception, we introduce assessment indefiniteness when calculating the evaluation accuracy of each measure to reduce the influence of errors of subjective perception. Experiments on various images are conducted to test the effectiveness of the proposed measure. The results have shown that the measure is useful in conjunction with the other objective measures to account for qualitative subjective perception.

Imagine a future in which computer-generated (CG) human characters become almost indistinguishable from real humans. It is not that far away. Real-time rendering of photorealistic facial animation is just around the corner. To prepare for this world in which CG content will be intermingled with the natural, there is a fundamental need to understand how human beings can preserve the awareness of what is real and what is not. We present a methodology for distinguishing real and CG human faces within the context of multimedia forensics, by exploiting the heart rate as the discriminating feature. The underlying idea is that humans always present a pulse signal that can be automatically extracted from a video sequence depicting their face, while this is not the case for virtual humans. Experimental validation shows the effectiveness of the proposed solution.

Depth image-based rendering (DIBR) has recently received significant attention as an efficient concept for generating virtual views. One of the main challenges in the view synthesis process is that holes will appear in the background (BG) of the virtual views because of occlusion. The existing methods either utilize spatial or temporal information from the original views and depth maps to fill the holes. However, the performance of the existing methods is seriously affected by the quality of the depth maps. Foreground (FG) texture will be wrongly introduced to the BG when the depth maps are not accurate enough. We develop a hole-filling approach for DIBR based on convolutional neural network, which shows reduced dependency on the quality of the depth map. More specifically, FG objects are detected on the RGB image by combining the Laplacian operator with a graph cut algorithm to avoid dependencies on the depth map quality. Furthermore, we learn an end-to-end mapping between the warped virtual view and the ground-truth view image with FG awareness. The learning-based approach exhibits excellent superiority in terms of saving time. The experimental results also show that the proposed approach outperforms current state-of-the-art techniques quantitatively and qualitatively.

Propagation-based methods have attracted more and more attention in saliency detection. We present a propagation-based saliency detection framework to overcome some drawbacks in existing methods. First, we construct an effective background-based map by defining a method to remove foreground noises in image boundaries. Second, a saliency propagation method is proposed to further refine the background-based map by introducing a feedback regulation mechanism into saliency propagation. Finally, an optimization method is developed to further enhance the contrast between salient object and background. Extensive experiments on five benchmark datasets are carried out, and experimental results show the superiority of the proposed method than other state-of-the-art methods.

Addressing the problem of iris certification security and unclear textures caused by defocused scanning, the constrained defocused state iris scanned by a low-end device is taken as the research object. A codeless mode certification algorithm for label statistical learning based on current optimal active feedback is proposed. The certification model is improved based on the convolutional neural network structure. Guided by statistical learning, the features from the same category irises are set to the feature label of one category, and the iris certification function is designed. In addition, a current optimal active feedback mechanism is proposed. The certification effects of the certification model will be actively reflected in real time based on the current training data and the accumulated certification data, and the user can adjust the model structure and parameters in real time, to maintain the current optimality of the certification model structure. In addition, a result mapping layer included a stealing response mechanism is set between the certification result and the user. The stealing behavior that occurs can be made meaningless by changing the result output mapping relationship. Under the algorithm prerequisites, the JLU iris library is used to conduct experiments, and the accuracy of the algorithm can reach more than 99%, which indicates that this algorithm can effectively improve the accuracy of constrained state defocused iris scans with a low-end device, the model can be adjusted dynamically, and stealing attack behavior can be responded to effectively.

To determine eggshell quality, many external factors, such as appearance and cleanliness, can be detected by human graders. However, the inspection of the internal factor of strength of the eggshell has proven difficult. Some eggs develop hairline cracks due to low strength; this regularly happens in the process of egg production, causing economic loss and exposure to Salmonella contamination. These developed hairline cracks may be detectable. Although the effectiveness is in doubt, potential crack development of the eggshell can be predicted by observing the translucent mottling patterns of the eggshell. Early discarding of eggs with severe mottling has thus become a common practice in egg production to prevent further crack development and cross-contamination. However, accurately predicting potential crack development based on eggshell mottling is a challenge due to the irregularities of the mottling pattern itself, with difference in shape, size, distribution pattern, and color inhomogeneity. A pretrained convolutional neural network (CNN) model based on AlexNet architecture was proposed to extract learned image features of the mottling pattern on eggshells. The classification results by the fine-tuned CNN model was then compared to the classifications by the human graders as ground truthing. The fine-tuned CNN model performed on par to the human graders, with 91.8% similarity in classification rate. Visualization and analysis on intermediate activation strengths in the fine-tuned CNN model confirmed the consistency and traceability of the classification results. Our work underscores the potential of the CNN-based feature extraction and classification on eggs based on highly inhomogeneous mottling patterns on eggshell.

Image fusion obtains a desired image by integrating the useful information of multiple input images. Most traditional fusion strategy is usually guided by image local contrast or variance, which cannot well represent visual discernable features of source images. And the undesirable seam effects or artifacts produced due to the inconsistency between fusion weight map and image content may severely degrade the visual quality of the fused images. An efficient image fusion method with structural saliency measure and content adaptive consistency verification was proposed. The fusion is implemented under the nonsubsampled contourlet transform (NSCT)-based image fusion framework. The low-frequency NSCT decomposition coefficients are fused with the weight map constructed by considering both structural saliency and visual uniqueness features and refined by spatial consistency with guide filter. The high-frequency NSCT decomposition coefficients are fused with structural saliency. The performances of the proposed method have been verified on several pairs of multifocus images, infrared-visible images, and multimodal medical images. Experimental results clearly demonstrate the superiority of the proposed algorithm compared with several existing state-of-the-art algorithms in terms of both visual and quantitative comparison.

Image registration is an indispensable operation in image processing. To solve the problem of nonlinear distortion which easily causes low registration property, a registration scheme including a nonlinear scale space, a new discrete orthogonal Gaussian–Krawtchouk (GK) descriptor vector, and a dual similarity measurement of distance-cosine is proposed. Nonlinear scale space is constructed to preserve a lot of important structures for increasing the quantity of the salient keypoints. Meanwhile, considering both blur stability and simple computation, multiscale Harris algorithm on the basis of nonlinear scale space is applied to extract the keypoints. Moreover, according to different scales, a descriptor vector based on multiscale GK invariant moments is presented and applied to the following image matching. Finally, a dual similarity measurement of the initial Euclidean distance and the fine space cosine is proposed for the purpose of enhancing the matching accuracy. The effectiveness of the proposed algorithm has been tested on the standard images with different types of variations. The experimental results show that the performance of feature extraction and the correct matching rate of the proposed scheme are superior to those of other classical methods. The proposed scheme has a better robustness and a higher registration accuracy.

We propose a synthetic image method named local-foreground generative adversarial networks for surface defect detection in the industry. The method comprises three contributions: First, in order to provide more training data, an algorithm that generates full synthetic images of defects is proposed. The method may blend the defect samples stored in the different mobile terminal into existing background images of the production environment in a natural way, accounting for both geometry and appearance. Second, the diversity of defects appearance is increased due to the use of deep convolutional generative adversarial networks only for local-foreground and the defect edge details that are preserved by introducing a guided filter in the process of image synthesis. Third, the image fusion method can adapt to the various production environment of different brightness and camera angles, which has strong adaptability and expansion ability. We also discuss the experimental results of synthetic data on an end-to-end detection system based on deep learning. As we know, YOLO detector trained on synthetic data achieving average precision of 95.2% on test dataset, which is 12.5% higher than heuristic data augmentation individually. Furthermore, it can process 23  images  /  s on 1080Ti GPU, which meets the requirements for real-time industrial inspection.

Generative adversarial network (GAN) for image super-resolution (SR) has attracted enormous interests in recent years. However, the GAN-based SR methods only use image discriminator to distinguish SR images and high-resolution (HR) images. Image discriminator fails to discriminate images accurately since image features cannot be fully expressed. We design a GAN-based SR framework GAN-IMC, which includes generator, image discriminator, morphological component discriminator, and color discriminator. The combination of multiple feature discriminators improves the accuracy of image discrimination. Adversarial training between the generator and multifeature discriminators forces SR images to converge with HR images in terms of data and features distribution. Moreover, in some cases, feature enhancement of feature-rich region is also worth considering. GAN-IMC is further optimized by weighted content loss (GAN-IMCW), which effectively restores and enhances feature-rich regions in SR images. The effectiveness and robustness of the proposed method are confirmed by extensive experiments on public datasets. Compared with state-of-the-art methods, the proposed method not only achieves competitive perceptual index and natural image quality evaluator values but also obtains pleasant visual perception in edge, texture, color, and feature-rich regions.

Recently, palmprint recognition has made huge progress and attracted the attention of more and more researchers. However, current research rarely involves open-set palmprint recognition. We proposed deep ensemble hashing (DEH) for open-set palmprint recognition. Based on the online gradient boosting model, we trained multiple learners in DEH, which focus on identifying different samples. In order to increase the diversity between learners, activation loss and adversarial loss were introduced. Through minimizing activation loss, the neurons of different learners restrained each other, and through adversarial loss, the optimal distance between the features extracted by different learners was obtained. Palmprint identification and verification experiments were performed on PolyU multispectral database and our self-built databases. The results show the effectiveness of DEH in deal with open-set palmprint recognition. Compared to baseline models, DEH increased the recognition accuracy by up to 6.67% and reduced the equal error rate by up to 3.48%.

We propose a three-dimensional video segmentation method using deep learning convolutional neural nets. The algorithm utilizes the local gradient computed at each pixel location together with the global boundary map acquired through deep learning methods to generate initial pixel groups by traversing from low to high gradient regions. A local clustering method is then employed to refine these initial pixel groups. The refined subvolumes in the homogeneous regions of video are selected as initial seeds and iteratively combined with adjacent groups based on intensity similarities. The volume growth is terminated at the color boundaries of the video. The oversegments obtained from the above steps are then merged hierarchically by a multivariate approach yielding a final segmentation map for each frame. The results show that our proposed methodology compares favorably well, on a qualitative and quantitative level, in segmentation quality and computational efficiency, with the latest state-of-the-art techniques utilizing the video segmentation benchmark dataset.

The moments constitute a well-known tool for image analysis and recognition tasks. The family of moments that has the most advantages is the discrete orthogonal moments. A set of these moments is the Hahn moments, as they have a great number of advantages in comparison with other sets of moments. The main disadvantage of moments, including Hahn moments, is the high computational cost, which is increased as higher order moments are involved in the computations, so the real-time analysis is hard to be done. We propose an effective approach for the computation of Hahn moments. The gray image is decomposed in a set of binary images that are named as bitplanes. The most significant bitplanes are represented using image block representation and their moments are computed fast using block techniques. The least significant binary images are substituted by a constant ideal image called “half-intensity” image, which has known Hahn moment values. The proposed method has low computational error, low computational complexity, and under certain conditions is able to achieve real-time processing rates.

We explore the use of an improved particle filter to detect and track stealth targets (STs) in noisy imagery with unknown detection profile. Recently, a cardinality-balanced multitarget multi-Bernoulli (CBMeMBer) of variational Bayesian (VB) approximation has been proposed for tracking multitargets with unknown measurement variances. However, VB-CBMeMBer filter has been implemented with a known detection profile, which is unsuitable for ST tracking. The Bayesian track-before-detect (TBD) is an efficient way to detect low observable targets. A new VB-CBMeMBer-TBD filter is proposed to cope with jointly unknown detection profile and measurement variances for ST tracking. Furthermore, a sequential Monte Carlo implementation is applied to estimate augmented kinematic ST state. The results show that the proposed filter is more accurate than other filters in cardinality estimation and less optimal subpattern assignment errors.