Salient Object Detection Method Based On Biological Visual Cognition Mechanism

Salient object detection aims at simulating the human visual attention mechansim to extract the most attractive areas of the scene.In recent years,research on saliency detection has become a hot topic in image processing and computer vision and has been widely utilized in many fields such as content transmission,image zoom,image segmentation,target recognition and so on.In this paper,by learning biological visual cognition mechanism,we apply it into saliency detection in the real world which address the effects of background noise,complicated design and low accurancy in traditional methods.The research contents are as follows:(1)Inspried by the biological visual parallel processing mechanism,a method of co-learning saliency maps with coupled channels and low-rank factorization is proposed.The basic idea is,that,firstly,the original image is divided into different superpixels and the pure background in image boundary template is extracted by the low rank matrix factorization method.Secondly,the original image is processed by two parallel paths and then obtain saliency maps based on “where” and “what” features.Finally,a fusion method is proposed to combine two feature maps,Based on this,the salieny map is further refined to obtain the final result.The experimental results shown that this method can effectively improve the detection accurancy,and can get better results in the ASD and ECSSD datasets.(2)According to the hierarchical cognitive mechanism of visual information,a detection method of RGB-D image saliency target detection based hierarchical cognitive mechanism is proposed.The saliency map acquisition can be divided into three stages,primary,intermediate and advanced acquisition.Firstly,a new method of RGB-D image superpixels segmentation based on depth information guidance is proposed.Secondly,we design primary saliency maps,intermediate saliency maps and advanced saliency maps based on visual contrast,graph cut and multi-scale fusion and semi-supervised ELM,respectively.Finally,the three maps are combined by a simple fusion method and then get the final map.The experiment results shown that the proposed method has achieved good results in visual effect and on PR curve,also achieves the most saient object detection results.(3)According to the sparseness of visual cognitive,a salient object detection method based on visual sparse cognition is proposed.The basic idea is that,Firstly,the image is divided into different superpixels,and all the superpixels are classified,the superpixel that does not belong to a certain class is used as the dictionary,so a salient object detect method based on global sparse reconstruction error is designed.Secondly,a salient object detection method based on local sparse reconstruction error is designed by locally constrained linear coding.Finally,a simple combination method is used to effectively integrate the global and local sparse reconstruction erroe method,and the final result is obtained.The validity of the method is verified on three public database,respectively.In addition,the method of this chapter is applied to the detection of ship target in high resolution SAR image.The experiment results shown that the algorithm of fusion of visual saliency has about 3%-5% increase compraed with traditonal ship detection algorithm.