Objective Shoeprints serve as vital evidence in forensic investigations, and determining how a massive amount of shoeprint images can be clustered automatically through long-term accumulation has become one of the urgent tasks of criminal technology. Unlike other image sets, the number of shoeprint categories is not only considerable but also unknown. Shoeprint images in the feature space are distributed inhomogeneously and sparsely, but the quantity of each class is low. For these reasons, most existing clustering algorithms cannot satisfactorily cluster shoeprints. In this study, an automatic clustering method is proposed to divide shoeprint sets effectively based on an analysis of the distribution of shoeprint images in the feature space. Method Through statistics on labeled shoeprint-image databases, we found that shoeprint sets of different patterns do not intersect, and a blank region, where no shoeprints exist, is present between every two classes. Blank regions are called margins in this paper. The core objective of the proposed algorithm is to determine the margins between classes and use them to divide a shoeprint set. The process involves the following steps:1) dividing the shoeprint set with monotonically increasing or descending thresholds, which are used to classify two shoeprint images into the same cluster; 2) searching for the cluster that does not change with K consecutive partitions; 3) outputting the stable cluster and removing the shoeprints belonging to the output stable cluster from the dataset; 4) choosing the next threshold and dividing the remaining dataset; 5) returning to step 2) until the remaining set is empty. Result Experimental results on two kinds of publicly available databases and one real shoeprint database which comprises 5792 images, have shown that the proposed algorithm outperforms state-of-the-art clustering algorithms on common clustering evaluation

Objective The LBP(local binary pattern) algorithm is sensitive to edge and noise. Thus, this study proposes a new algorithm called uniform local mean pattern (ULMP). Considering the complementarity of global and local features on recognition, this study proposes a face recognition method based on ULMP description and double weighted decision fusion for classification. Method First, we use the ULMP algorithm to derive the code diagram of the entire image. The concrete steps to obtain the eight binary codes are implemented by comparing the eight average pixels with the center pixel. Each of the eight values is obtained by computing the average pixels of the eight directions, three horizontal directions, three vertical directions, and two diagonal directions. Each binary code is multiplied by the corresponding weight coefficient and then added to derive the ULMP coding value of the center pixel and the code pattern of the entire image. Then, the code diagram is divided into equal sub-blocks and each sub-block histogram is assessed to determine the local texture features. The global texture feature is obtained by connecting the histogram of different sub-block features. To emphasize the importance of different sub-blocks in the final recognition, this study introduces the cloud model and structure-based classifiers by constructing sub-image sets to obtain the weight of each sub-block. In the testing phase, each block of the statistical characteristics of a test sample is combined with the BP neural network to determine the posterior probability of each category. We use the weights calculated by the cloud model fused with the linear weighted decision to derive the local classification results. After obtaining the results of local and global classification, we conducted weighted integration to obtain the final recognition results. Result The experimental results on the ORL and Yale face database show that the ULMP exhibits good recognition performance. The average recognition rate is 95.9% on the ORL database with five test samples. The proposed approach increases the recognition rate of LBP, MCT(modified census transform), LGP(local gradient patterns), ULBP, and CSLBP by 11.3%, 10.6%, 9.5%, 8.9%, and 3.9%, respectively. The average recognition rate is 97.4% on the Yale database with five test samples. The proposed approach increases the recognition rate of LBP, MCT, LGP, ULBP, and CSLBP by 18.9%, 17.7%, 17.1%,

Objective All traditional example-based super-resolution methods adopt image-gradient features for low-resolution images and thus, these methods are unable to characterize the low-resolution space satisfactorily. To address this issue, this paper proposes a novel unified framework for image super-resolution that effectively combines example-based method with deep learning models. Method The proposed method consists of three main stages:low- and high-resolution similarity-learning, high-resolution patch-dictionary-learning, and high-resolution patch-generating stages. At the first stage, two different convolutional neural networks are proposed for learning a novel similarity metric between high- and low-resolution image patches. At the second stage, the high-resolution patch dictionaries are learned from training sets. At the last stage, the high-resolution patches are generated based on learned similarities between the input low-resolution patch and the atoms in the high-resolution patch dictionary. Result Experimental results on several commonly adopted datasets show that the proposed two-channel model quantitatively and qualitatively achieves improved performance compared with other methods. Conclusion The proposed two-channel model can preserve more detailed information and reduce ringing artifacts in the resulting images.