Multi-Channel Hierarchical Feature Extraction For Image Recognition

Image recognition is a hot research subject in some research fields such as pattern matching, computer vision and machine learning and so on. Through years of research and development, there are many mature technologies that have been developed in the field of image recognition and are widely applied to various fields, for example, remote sensing, aerial photography, license plate recognition and monitoring, and the like. In the complex real world, many images photographed by people contain information of multiple dimensions, for example, size, perspective, light condition and scene layout, etc. In the field of image recognition, if we focus primarily on the recognition of an object in the image, information such as light, scene and the like in the image will become interference information. With too much interference information, extracting desired effective features in the images will be particularly difficult. Thus, it has become an important issue in the computer vision as to how to extract the effective features of the images without being influenced by the interference information.With respect to the above-mentioned issues, this paper provides a hierarchical feature extraction algorithm under multi-channel, which eliminates the influence on the image feature extraction by the interference information and improves the algorithm recognition rate. The major work in this paper is listed as follows:(1) It has been found that an over-fitting phenomenon easily occurs in an existing dictionary learning algorithm K-SVD. Therefore, a coherence parameter is introduced in the K-SVD algorithm, thereby reducing the degree of correlation of atoms in the dictionary obtained from learning and enhancing the expression ability of the dictionary. The experiment result shows that the recognition rate has increased by 2%-4% compared to that before optimization and the efficiency of the mixed programming of Matlab and C-MEX is 23 times larger than that of Matlab.(2) The most popular sparse coding algorithm is analyzed, and advantages and disadvantages of a matching pursuit algorithm and an orthogonal matching pursuit algorithm are compared. Thereafter, spatial pyramid pooling in the convolutional network is applied to the sparse coding so as to reduce the computation complexity.(3) A multi-channel method is designed to perform feature extraction, that is, images are divided into different sizes, each dividing scheme (channel) obtains feature vectors of the image after multilevel feature extraction. Then a linear SVM-trained model is used for the feature vectors of different channels for recognition. The experimental result shows that the result after adding the multi-channel and sparse coding is at least 4% higher than the recognition rate of other feature extraction algorithms.