The paper deals with JPEG adaptive lossy compression of color images formed by digital cameras. Adaptation to noise
characteristics and blur estimated for each given image is carried out. The dominant factor degrading image quality is
determined in a blind manner. Characteristics of this dominant factor are then estimated. Finally, a scaling factor that
determines quantization steps for default JPEG table is adaptively set (selected). Within this general framework, two
possible strategies are considered. A first one presumes blind estimation for an image after all operations in digital
image processing chain just before compressing a given raster image. A second strategy is based on prediction of noise
and blur parameters from analysis of RAW image under quite general assumptions concerning characteristics
parameters of transformations an image will be subject to at further processing stages. The advantages of both strategies
are discussed. The first strategy provides more accurate estimation and larger benefit in image compression ratio (CR)
compared to super-high quality (SHQ) mode. However, it is more complicated and requires more resources. The second
strategy is simpler but less beneficial. The proposed approaches are tested for quite many real life color images acquired
by digital cameras and shown to provide more than two time increase of average CR compared to SHQ mode without
introducing visible distortions with respect to SHQ compressed images.
In the camera manufacturing, special methods are needed to detect blemishes occurring on the camera sensor pixels. A blemish is referred as a region of pixels in the camera sensor that are somewhat darker than the background. The blemishes are difficult to detect accurately, but on the other hand, they cause a significant reduction in camera quality. We present a novel filtering method for the blemish detection. The method is based on image scaling, filtering, and difference image calculation that is very fast and accurate in the detection of blemishes. In addition, the algorithm can cope with unprocessed raw image data, in which various distortions, such as noise and vignetting, can be present.
Classifier combinations can be used to improve the accuracy of demanding image classification tasks. Using combined classifiers, nonhomogenous images with noisy and overlapping feature distributions can be accurately classified. This can be made by classifying each visual descriptor first individually and combining the separate classification results in a final classification. We present an approach to combine classifiers in image classification. In this method, the probability distributions provided by separate base classifiers are combined into a classification probability vector (CPV) that is used as a feature vector in the final classification. The proposed classifier combination strategy is applied to the classification of natural rock images. The results show that the proposed method outperforms other commonly used probability-based classifier combination strategies in the classification of rock images.
In image classification, the common texture-based methods are based on image gray levels. However, the use of color information improves the classification accuracy of the colored textures. In this paper, we extract texture features from the natural rock images that are used in bedrock investigations. A Gaussian bandpass filtering is applied to the color channels of the images in RGB and HSI color spaces using different scales. The obtained feature vectors are low dimensional, which make the methods computationally effective. The results show that using combinations of different color channels, the classification accuracy can be significantly improved.
The use of image retrieval and classification has several applications in industrial imaging systems, which typically use large image archives. In these applications, the matter of computational efficiency is essential and therefore compact visual descriptors are necessary to describe image content. A novel approach to contour-based shape description using wavelet transform combined with Fourier transform is presented. The proposed method outperforms ordinary Fourier descriptors in the retrieval of complicated industrial shapes without increasing descriptor dimensionality.
Clustering of the texture images is a demanding part of multimedia database mining. Most of the natural textures are non-homogenous in terms of color and textural properties. In many cases, there is a need for a system that is able to divide the non-homogenous texture images into visually similar clusters. In this paper, we introduce a new method for this purpose. In our clustering technique, the texture images are ordered into a queue based on their visual similarity. Based on this queue, similar texture images can be selected. In similarity evaluation, we use feature distributions that are based on the color and texture properties of the sample images. Color correlogram is a distribution that has proved to be effective in characterization of color and texture properties of the non-homogenous texture images. Correlogram is based on the co-occurrence matrix, which is a statistical tool in texture analysis. In this work, we use gray level and hue correlograms in the characterization of the colored texture. The similarity between the distributions is measured using several different distance measures. The queue of texture images is formed based on the distances between the samples. In this paper, we use a test set which contains non-homogenous texture images of ornamental stones.
Clustering of the images stored in a large database is one of the basic tasks in image database mining. In this paper we present a clustering method for an industrial imaging application. This application is a defect detection system that is used in paper industry. The system produces gray level images from the defects that occur at the paper surface and it stores them into an image database. These defects are caused by different reasons, and it is important to associate the defect causes with different types of defect images. In the clustering procedure presented in this paper, the image database is indexed using certain distinguishing features extracted from the database images. The clustering is made using an algorithm, which is based on the k-nearest neighbor classifier. Using this algorithm, arbitrarily shaped clusters can be formed in the feature space. The algorithm is applied to the database images in hierarchical way, and therefore it is possible to use several different feature spaces in the clustering procedure. The images in the obtained clusters are associated with the real defect causes in the industrial process. The experimental results show that the clusters agree well with the traditional classification of the defects.