| Image Target Recognition technology is an important research field of image understanding and computer vision, with its wide range of application, it attracts more and more attention of scholars. In recent years, inspired by the human visual system research, Deep Learning had been proposed, and it quickly became a hot spot. Deep Learning model with a multi-layer structure can overcome the shortcomings of the traditional methods effectively, but the Deep Learning model is generally abstract, and it is difficult to understand. HMAX model is a network of four layers, and it can complete the classification for a particular class. Its structure is the same as the convolutional neural network, and in fact, it is considered as a convolutional neural network which joined with the prior information of human visual system.The main work and achievements of this paper are as follows:1, By studying the directional characteristics of Gabor filters, a S1 layer sparsing method is proposed to reduce the time-consuming of S1 layer. Gabor filters have directional selectivity, when the direction of filter is parallel to the tangential direction towards the edge of the image, the filter will give the maximum response. According to this, when generating the S1 layer, the Gabor filter only convolute with the pixels whose tangential direction is the same as the Gabor filter. Experimental results show that such an operation can reduce the time-consuming effectively but without losing image information of C1 layer.2, By studying the extracting techniques of feature point, a new SIFT feature extraction algorithm is proposed to resolve the redundancy of C1 patches. The extraction problem of C1 patches can be attributed to a extraction problem of feature point. And the new SIFT feature extraction algorithm can effectively remove the false SIFT feature points. Experimental results show that the obtained feature point mainly drops in target area.3. By studying the techniques of feature selection, a new feature selection algorithm based on Relief and SVM-RFE is proposed to reduce the vector dimensions of C2 layers. The new feature selection can extract informative features while eliminate the redundant features, irrelevant features, noisy features effectively. Experimental results show that the new algorithm helps to improve classification accuracy, better than both Relief algorithm and SVM-RFE algorithm.4, Apply the improved HMAX models to Caltech101 database. The experimental results show that improved HMAX has a better performance. |
