Convolutional Neural Networks (CNNs) have recently demonstrated outstanding performance in image retrieval tasks. Local convolutional features extracted by CNNs, in particular, show exceptional capability in discrimination. Recent research in this field has concentrated on pooling methods that incorporate local features into global features and assess the global similarity of two images. However, the pooling methods sacrifice the image's local region information and spatial relationships, which are precisely known as the keys to the robustness against occlusion and viewpoint changes. In this paper, instead of pooling methods, we propose an alternative method based on local similarity, determined by directly using local convolutional features. Specifically, we first define three forms of local similarity tensors (LSTs), which take into account information about local regions as well as spatial relationships between them. We then construct a similarity CNN model (SCNN) based on LSTs to assess the similarity between the query and gallery images. The ideal configuration of our method is sought through thorough experiments from three perspectives: local region size, local region content, and spatial relationships between local regions. The experimental results on a modified open dataset (where query images are limited to occluded ones) confirm that the proposed method outperforms the pooling methods because of robustness enhancement. Furthermore, testing on three public retrieval datasets shows that combining LSTs with conventional pooling methods achieves the best results.

In this paper, impulse noise removal for digital images is handled. It is well-known that switching-type processing is effective for the impulse noise removal. In the process, noise-corrupted pixels are first detected, and then, filtering is applied to the detected pixels. This switching process prevents distorting original signals. A noise detector is of course important in the process, a filter for pixel value restoration is also important to obtain excellent results. The authors have proposed a local similarity-based filter (LSF). It utilizes local similarity in a digital image and its capability against restoration of orderly regions has shown in the previous paper. In this paper, first, further experiments are carried out and properties of the LSF are revealed. Although LSF is inferior to an existing filter when disorderly regions are processed and evaluated by the peak signal-to-noise ratio, its outputs are subjectively adequate even in the case. If noise positions are correctly detected, capability of the LSF is guaranteed. On the other hand, some errors may occur in actual noise detection. In that case, LSF sometimes fails to restoration. After properties are examined, we propose two effective extensions to the LSF. First one is for computational cost reduction and another is for color image processing. The original LSF is very time consuming, and in this paper, computational cost reduction is realized introducing a search area. Second proposal is the vector LSF (VLSF) for color images. Although color images can be processed using a filter, which is for monochrome images, to each color component, it sometimes causes color drift. Hence vector processing has been investigated so far. However, existing vector filters do not excel in preservation of orderly pattern although color drift is suppressed. Our proposed VLSF is superior both in orderly pattern preservation and color drift suppression. Effectiveness of the proposed extensions to LSF is verified through experiments.

To reconstruct low-resolution facial photographs which are in focus and without motion blur, a novel algorithm based on local similarity preserving is proposed. It is based on the theories of local manifold learning. The innovations of the new method include mixing point-based entropy and Euclidian distance to search for the nearest points, adding point-to-patch degradation model to restrict the linear weights and compensating the fusing patch to keep energy coherence. The compensation reduces the algorithm dependence on training sets and keeps the luminance of reconstruction constant. Experiments show that our method can effectively reconstruct 1612 images with the magnification of 88 and the 3224 facial photographs in focus and without motion blur.

In reconstruction-based super resolution, a high-resolution image is estimated using multiple low-resolution images with sub-pixel misalignments. Therefore, when only one low-resolution image is available, it is generally difficult to obtain a favorable image. This letter proposes a method for overcoming this difficulty for single- image super resolution. In our method, after interpolating pixel values at sub-pixel locations on a patch-by-patch basis by support vector regression, in which learning samples are collected within the given image based on local similarities, we solve the regularized reconstruction problem with a sufficient number of constraints. Evaluation experiments were performed for artificial and natural images, and the obtained high-resolution images indicate the high-frequency components favorably along with improved PSNRs.