Research Of Object Detection Based On Visual Saliency

The explosive growth of images and videos has brought great convenience for human's education,entertainment and consumption,but has caused big challenges for visual information processing such as analysis and application of image and video.In order to put the limited computing resources on visual information processing and extract the valuable content from massive information,the researchers have carried out the study of visual saliency detection through human visual attention.Saliency detection can be divided into two research areas: eye fixation and salient object detection,and this thesis focuses on the study of salient object detection.The recently proposed salient object detection models have created many valuable features and priors,however,these features and priors just keep a good results on specific scene images.Thus,this thesis focuses on the defects of these features and priors such as objectness,compactness,contrast and boundary connectivity,modifies the feature fusion strategies and proposes five models to enhance their adaptability to different scene images and improve the performance of salient object detection.The main work and contribution of this thesis are as follows:(1)The model of improved saliency optimization based on superpixel-wised objectness and boundary connectivity is proposed.Objectness can locate the rough position of salient object in image,but cannot show the boundary of the targets.However,the superpixel segmentation can identify the edges of object and reduce the computation.So,the superpixel-wised objectness is porposed by combining these two parts.Boundary connectivity can better distinguish target from background,so the saliency map is obtained by multiplying the modified objectness and boundary connectivity.Finally,a better salient detection result is gotten by optimiting the saliency map with center prior.This model can reduce the deficiency of original objectness,and improve the accuracy of saliency detection.However,when the target is large or composed of multiple parts,the results of this model is poor.(2)The model of saliency detection by hierarchically integrating compactness,contrast and boundary connectivity is proposed.In order to make the large or multi-part targets can be effectively detected in images,the model ranks the similarity matrix and obtains the weighted compactness and contrast feautres based on the matrix.The size and shape of objects in images are different,so that the saliency result is poor when the model is performed on single segmentation scale.Therefore,given the computational efficiency,the model obtains the final saliency maps by integrating the weighted compactness,contrast and boundary connectivity based on four different segmentation scales,which further improve the saliency results.The model is effective for images with simple target and background,but when the background becomes complex,it is easy to cause false background detection.(3)The model of salient object detection based on compactness and foreground connectivity is proposed.In order to further improve the performance on images with complex background,with the similarity matrix as weight,the spatial contrast based compactness is obtained by weighting the spatial distances between image blocks.The center contrast based compactness is obtained by weighting the distances from image blocks to image center.Then,the manifold ranking based compactness is achieved by fusing and ranking these two contrast based compactness.Furthermore,the foreground seeds are obtained by the fused compactness,and the foreground connectivity is proposed by ranking these foreground seeds.The final salient results are obtained by integrating the fused compactness and foreground connectivity at different scales.The model can highlight complete targets and suppress the influence from complex background.However,when the object is similar to background,the model will detect these background regions as salient object.(4)The model of salient object detection via manifold ranking on multi-feature based graphs is proposed.To enhance the ability to distinguish object from similar background,the model in this section chooses multi-feature to construct the graphs.Then,based on these graphs,the regions of image boundary and boundary connectivity are ranked respectively to get two different salient results.The final salient map is obtained by integrating these two different salient results at different scales.Compared with the manifold ranking results based on single feature,the results of this model can effectively divide object from background,and enhance the performance of the model.However,this model can generate artifacts around the detected salient object.(5)The model of salient object detection based on manifold ranking and co-connectivity is proposed.To reduce the artifacts around the salient object,the model in this part uses multifeature to calculate the weights,and modifies the connections by threshold to get new graphs.And based on these new graphs,the model ranks the boundary regions and integrates them to obtain the boundary based saliency maps.Simultaneously,the co-connectivity based salient map is obtained by incorporating boundary connectivity and foreground connectivity.The final salient result is obtained by fusing these two salient maps at different segmentation scales.The model can reduce the artifacts around the target and improve the performance effectively by selecting the edge connection of graphs.