Research On Biological Visual Features Extraction In Images Combining Sparse Coding

way to extract better image features is always the research focus. Human visual system has highly evolved, which can easily fulfill the task of image classification and is superior to current any computer visual system. With the continuous development of neuroscience, the understanding of human visual perception system has continued to deepen. According to the research in neuroscience, how to model the information processing of human visual system and establish the biological visual feature extraction model has always been a challenging and attractive research area.Inspired by the biological vision, recently Serre and Poggio et al. have proposed a quantitative hierarchical model of object recognition based on the theory of the feedforward path of ventral stream of visual cortex, which has been called Standard Model of Visual Cortex(ST model), also called HMAX model. This model is composed by four layers which are called S1 layer,C1 layer,S2 layer and C2 layer. The feature extracted by the ST model is called the Standard Model Feature(SMF), which has the invariance to scale and position, invariance to rotation could be obtained via the training set by introducing rotated versions of the original input. ST model acquires the redundant intermediate complexity feature set which is also called feature dictionary of intermediate layer by selecting patches from C1 layer randomly. However, this method of construting feature dictionary is arbitrary and lack of biological basis, so it should be modified to be more biologically plausible.Neuroscience research results have shown that sparse coding is the primary form of image transformation in the visual system. The response of V1 neurons to the nature images is sparse. The V4 neurons represent the visual information via sparse coding. These results provide biological basis for bringing sparse coding into the ST model.This thesis introduces sparse coding into the ST model, proposes a method of image biological visual feature extraction combining sparse coding, which is called Sparse Coding ST Model. This model acquires the feature dictionary of intermediate layer via K-SVD algorithm and calculates the S2 feature via sparse coding, then the C2 feature is obtained from S2 layer which is called Sparse Coding SMF(SCSMF).The SMF and SCSMF are compared via image classification experiments. The results show that SCSMF is better than SMF, which can extract the biological visual features in images effectively.