Image Object Recogniyion Based On Visual Perception Mechanism

Image object recognition is an important research direction in computer visual field and its main target is to determine what objects are in the image and where they are. Object recognition has significant application value in the fields of remote sensing image recognition, intelligent transportation systems, military target recognition etc. In this paper, the visual mechanism is applied to object recognition. Also, the bottom-up visual saliency, top-down visual saliency and the fusion of both, as well as the feature aggregation and image description are studied.Firstly, the human visual attention mechanism can quickly locate on the interest regions, obtain the useful information and exclude the irrelevant information. Accordingly, the human visual perception mechanism is attempted to simulate and an object recognition algorithm based on the bag-of-word model of the bottom-up visual saliency regions is proposed in this paper. Then, the bottom-up visual saliency map of the image is calculated, and the OSTU image segmentation is used to get the most salient regions. After that, the salient region features are extracted, the image description is obtained, and the classification and recognition is eventually achieved with the SVM method.Secondly, since the spatial pyramid matching pooling often produces inconsistent spatial representation when the interest object appears in different positions of the image. This paper presents a salient-region based centric pooling method. Then, the sliding window method is used to obtain the salient object region that can be taken as the foreground region. Also the features from the foreground and background regions are extracted respectively to form the foreground and background spatial pyramid image description. This method achieves the ideal result.Finally, an object recognition method based on the combination of the bottom-up visual saliency and the top-down visual saliency is presented. Due to the nature of the data driven, the bottom-up visual saliency often responses to the irrelevant background, and misses the interesting object information. A new method to combine the bottom-up and top-down visual saliency is therefore introduced in this paper. This method fuses the bottom-up bio-inspired visual saliency information and the top-down prior information. And more reasonable visual saliency can be obtained, so as to improve the recognition rate.