A Novel Class-Specific Deep Convolutional Feature Selection Method for Image Representation
Tian Zhao, Ting Shu, Huan Qi Zhao · 2023
Image representation is a critical problem in computer vision, which can improve the result of image classification, object detection, visual tracking, and so on. The image feature can be divided into two categories, the handcrafted feature and the deep convolutional feature (DCF). Compared with the handcrafted feature, the DCF based on deep convolutional neural networks is more effective and efficient. Therefore, more and more researchers utilized DCF in various computer vision tasks. However, most existing works directly use or simply reduce the dimension of DCF, which may induce performance degradation in practical applications due to its high degree of redundancy. To tackle this problem, we propose a novel DCF selection method named class-specific deep convolutional feature selection (CS-DCFS) for effective image representation. As various classes are distinguished from others based on different feature subsets, the CS-DCFS selects a specific feature subset for each class from the DCF candidate set. A modified genetic algorithm that uses one of the two mutation operations is proposed to select the feature, two designed mutation operations can increase the generating speed of the population. Furthermore, class-separability is adopted to measure the selected feature subset, which is a powerful and effective measurement of feature discrimination. Experimental results have shown the effectiveness and efficiency of the proposed method on the image classification task in four challenging benchmark datasets.