Search

In this paper, we present a system for segmentation and classification of breast cancer ROI images by integrating the idea of hierarchical processing of segmentation and classification tasks. The system is composed of a segmentation module and a GCN module, where the GCN module is designed to improve the performance of the classification result. The segmentation module is used to obtain the segmentation masks of the image patches of the ROI image, which are spliced to generate the segmenion result of the ROI image. The GCN module is used to capture the spatial and semantic dependencies among image patches of the ROI image, by constructing the graph using the segmentation masks of the image patches. Based on the learned features by the GCN module, the classification result of the ROI image can be obtained. Experimental results on the grand challenge on BreAst Cancer Histology images (BACH) 2018 dataset [17] show that, the proposed segmentation and classification method outperforms the winner of BACH 2018 significantly, which demonstrates the effectiveness of the proposed method.

There are a wide range of applications for scene text detection and recognition due to the increasing popularity of portable digital devices. However, large-scale evaluation benchmark with multilingual and multi-oriented texts is still slow to occur to facilitate the research on scene text detection and recognition. In this paper, a large-scale and well-annotated scene text dataset, namely STV2k, is presented, which can be used for scene text detection as well as scene text recognition. Since all the images are collected from streets by smart phone, the textual scenes are rich of variations in layouts, color, fonts and backgrounds. Two state-of-the-art algorithms for scene text recognition are tested on this newly built dataset. The preliminary experiments demonstrate how challenging the scene text recognition is in real scenario.

Scene character recognition is capturing increasing interests due to the renewed interests in scene text recognition. For scene character recognition, feature representation is an important issue, which has been pursued in recent years. In this paper, we propose to use PCANet (principal component analysis network) to learn character features for scene character recognition. PCANet is proposed in [1], and it is a kind of deep learning framework by cascading PCA as in convolution neural network. In this paper, we apply PCANet to scene character recognition by customizing the architectures of PCANet to characters. The proposed method achieves promising performance on the Chars74K-15 dataset (achieving an accuracy of 64%) and the ICDAR03-CH dataset (achieving an accuracy of 75%), demonstrating the effectiveness of PCANet in scene character recognition.

Context has been successfully applied to 2D image processing in the tasks of object detection, object recognition, etc. However, it is usually ignored in 3D object tracking. In this paper, we propose a stereo context-based object tracker that explores 3D context information for collaborative tracking in binocular video streams. The target is located by voting from companions, which are keypoints that have strong 3D motion correlations to the target. The companions are obtained by using spatial-temporal constraints in 3D space. Quantitative experimental results on challenging video sequences demonstrate the robustness and superior performance of the proposed tracker.