Visual Saliency Modeling Based On Object Perception

With the development of intelligent information technology,images have been integrated into all aspects of people's lives.While bringing great convenience to people,a large number of image information also poses a great challenge to related processing technology.As one of the representative problems in the field of computer vision,saliency detection aims to efficiently and accurately detect the region of interest in the scene by simulating the human bottom-up and top-down visual attention mechanisms.It is widely used in various computer vision tasks,industrial production defect detection and so on.Based on the theory of visual attention mechanism,this paper studies the image saliency detection algorithm in different periods.On this basis,two image saliency detection models are established.The main work includes:1)We proposed a saliency detection algorithm combining multiple prior and sparse reconstruction.The algorithm improves the traditional background prior knowledge and extracts more reliable background template,which effectively solves the problem of salient region contacting with the image boundary.From the perspective of background and foreground,on the one hand,the algorithm calculates the regional contrast by color feature and spatial information,and the saliency map based on the reliable background template is obtained.On the other hand,the sparse error reconstruction is performed with the reliable background template as the dictionary of sparse representation.Thus,the salient information can be recovered,and the Gaussian model of object bias is used to enhance salient information to obtain a saliency map.Finally,two saliency maps are fused by fusion strategy.The average accuracy of the algorithm on MSRA1000 dataset is 89.92%,the average recall rate is 86.14%,and F-measure is 89.02%;the average accuracy on ECSSD dataset is 78.92%,the average recall rate is 70.84%,and F-measure is 76.89%.Experimental results show that the proposed algorithm can effectively suppress background information while detecting salient object.2)We introduce an image saliency detection algorithm guided by object perception.The algorithm constructs a fully convolution neural network to effectively learn the rich high-level semantic information in the image.We designed the deconvolution operation to solve the input-output resolution mismatch problem,and obtain the pixel-level salient object region through network training.Then,aiming at the problem of salient object edge blur,we use the results of object perception to guide the manifold ranking at multiple scales,and complete the spread of the object edge saliency.We performed experiments on three more complex datasets.Compared with the most advanced of the nine comparison methods,the F-measure value of this algorithm increased by 3.0% on the ECSSD dataset,increased by 9.2% on the DUT-OMRON dataset,and increased by 1.5% on the SED2 dataset.Experimental results show that the saliency object detected by this algorithm is more complete and the detection result is more robust in complex scenes.