Multi-channel biomimetic visual transformation for object feature extraction and recognition of complex scenes

摘要

Object recognition occurs accurately with human visual neural mechanism despite in different complex background interference. For computer system, it is still a challenging work of object recognition and classification. Recently, many methods for object recognition based on human visual perception mechanism are presented. However, most methods cannot achieve a better recognition accuracy when object images are corrupted by some background interferences. Therefore, it is necessary to propose a method for object recognition of complex scene. Inspired by biomimetic visual mechanism and visual memory, a multi-channel biomimetic visual transformation (MCBVT) is proposed in this paper. MCBVT involves three channels. Firstly, some algorithms including orientation edge detection (OED), local spatial frequency detection (LSFD) and weighted centroid coordinate calculation are adopted for two stage's visual memory maps creations during the first channel, where some visual memory points are stored in memory map. Secondly, an object hitting map (OHM) is built in the second channel and the OHM is an edge image without background interference. After that, the first stage's visual memory hitting map is obtained through execute back-tracking second stage's visual memory map. Furthermore, an OHM is constructed through back-tracking with common memory points in first stage's visual memory map and first stage's visual memory hitting map. Thirdly, the OED and LSFD algorithms are conducted to extract a feature map of OHM in the third channel. Consequently, the final feature map is reshaped into a feature vector, which is used for object recognition. Additionally, several image database experiments are implemented, the recognition accuracy for alphanumeric, MPEG-7 and GTSRB database are 93.33%, 91.33 and 90% respectively. Moreover, same object images in different backgrounds share with highly similar feature maps. On the contrary, different object images with complex backgrounds through MCBVT show different feature maps. The experiments reveal a better selectivity and invariance of MCBVT features. In summary, the proposed MCBVT provides a new framework of feature extraction. Background interference of object image is eliminated through the first and second channel, which is a new method for background noise reduction. Meanwhile, the results show that the proposed MCBVT method is better than other feature extraction methods. The contributions of this paper is significant in computational intelligence for the further work.