Object Detection Based On Mid-level Features

As a core research topic in computer vision, object detection is widely employed in robot environment perception, pedestrian detection, video analysis, and image retrieval etc. Since there are inevitable influences of illumination, view, scale, and non-rigid changes when capturing image, the appearance features of objects will greatly change which leads to challenge for detection tasks. Feature extraction and representation is the key step of object detection. Low-level feature representations have had great success in visual recognition problems, yet there has been a growing body of work suggesting that the traditional approach of using only low-level features may be insufficient. Different to low-level features, mid-level features, a structured image description(e.g. parts), have attracted more and more attention since they capture semantic concepts. Significant performance gains can be achieved by introducing mid-level features beyond the plain visual cues offered by low level features. However, how to efficiently extract mid-level features is still a challenge problem. In this thesis, inspired by biological vision system, object detection approaches based on mid-level feature presentations have been investigated. The main contents are as follows:(1) An Exemplar SVM based object detection method using visual attention mechanism is proposed. Firstly, after extracting patches of images, the Normed Gradient(NG) features of the patches are extracted to separate foregrounds and backgrounds. The saliency regions of images are also obtained using a SVM classifier. Then a learning method based on exemplar SVM is proposed for fine object detection. Experimental results over different datasets show that the proposed method performs better than other counterparts in terms of detection precision.(2) An object detection method based on Edge structure Similarity Region Convolutional Neutral Networks(ESRCNN) is presented. Firstly, an edge structure similarity strategy is applied to select more efficient mid-level image patches from the candidates. Then using the selected image patches as input, the Convolutional Neutral Networks(CNN) is applied to extract features. Finally, a SVM classifier is utilized for detection. Experimental results show that the proposed method can obtain better performances than other counterparts.(3) An object detection method based on multi-channel hierarchical feature representation is proposed. Firstly, the Simple Linear Iterative Clustering(SLIC) algorithm is employed to generate superpixels, and some representative and discriminative image patches are discovered according to entropy-rank criterion. Then, the hierarchical convolution features are generated using convolution of selected patches and region proposals obtained by a selective search method. Next, the hierarchical convolution features are combined with the features extracted by the CNN channel to describe the characteristics of objects. Finally, the SVM is applied to classify the features. Experimental results demonstrate that the proposed method performs better than other counterparts.