Research On Saliency Detection Based On Visual Perception

The development of artificial intelligence technology has not only changed people's lifestyles,but also promoted the development of related fields,especially in the field of computer vision,including processing methods and algorithm performance,which has been greatly promoted.Nowadays,the research of vision algorithms has shifted from traditional manual design to the adaptive learning process based on big data analysis and reasoning.Taking the complexity of the application into account,the data involved has also changed from the early low-resolution,simple two-dimensional image data to large-scale multi-scene,high-resolution three-dimensional geometric data.How to improve the performance of the algorithm has become an urgent problem.Saliency detection by simulating the human visual system is an important research direction in the field of computer vision,which has attracted more and more attention.The goal of saliency detection is to allocate the limited computing resources to the regions where people are most interested in the scene to achieve the most effective resource allocation.When the significant region is obtained,the subsequent analysis and identification will be focused on.This not only provides effective information for the subsequent tasks,reduces the workload,but also greatly improves the performance of subsequent algorithms.For different data types,including two-dimensional image data and three-dimensional geometric data,different saliency detection algorithms are proposed in this paper.The main work is as follows:(1)Image salient object detection based on commute-time distance.For the problem of image saliency detection,a method for measuring regional saliency based on commute-time distance is proposed.Commute-time distance comprehensively considers all feasible paths between nodes,rather than a single path,which is a robust measurement method.Firstly,the classic K-means clustering algorithm is used to filter out the more confident background seed points of the image boundary,and the background prior map is constructed;Secondly,harris corner is used to construct the convex hull surrounding the salient object,and the rough position of the salient object is located.The reliable foreground seed points in the convex hull are induced by the background seed points,and the improved convex hull prior map is constructed.Finally,the final saliency map is generated by fusing two prior maps.When the inter-regional feature comparison used in this algorithm obtains the two initial saliency maps,both novel and robust commute-time distances are applied.Experimental results show that commute-time distance can more accurately measure the similarity of intervals and avoid the influence of background noise.(2)Mesh saliency detection based on entropy.For the problem of mesh saliency detection,a mesh saliency detection method based on information entropy is proposed.This method mainly uses the local contrast mechanism to measure the salient region.Firstly,the normal vector of the mesh vertex is calculated and selected as the descriptor of the vertex;Secondly,the probability distribution of the normal vector in the neighborhood of each vertex is calculated based on histogram method;Finally,the entropy of the normal vector in the neighborhood of each vertex is calculated,namely the saliency of the vertex.The entropy of the normal vector is used to describe the degree of change in the local neighborhood,If the entropy value is large,the local change is large,and the normal vector is chaotic,which means a salient region;if the entropy value is small,the local change is small,and the normal vector is in order,which means a non-salient region.The algorithm only needs to calculate the entropy of the normal vector,so the algorithm has certain advantages in running time.In addition,the saliency results are applied to mesh smoothing and mesh simplification,and good experimental results are also obtained.(3)3D grasp saliency analysis via deep shape correspondence.For the problem of saliency analysis of 3D shape grasping,a deep learning framework with data enhancement strategy as the pre-task is proposed.First of all,for supervised deep learning methods,the training data coverage area will greatly affect the model training,and the annotation of the captured data is extremely complicated.Therefore,a semantic feature matching based pre task is proposed,which transforms the dataset generation problem into a non-rigid shape correspondence,and achieve significant and effective migration through semantic correspondence between threedimensional shapes,thereby obtaining a larger-scale dataset.Secondly,considering that the three-dimensional shape capture task is not only related to the local features of the model,but also closely related to the overall structure of the model.therefore,the global features of the model are integrated into the network training to enhance the prediction performance of the network.Experiments show that through the use of pre-tasks,the diversity and scale of training data can be effectively improved.Training the network based on the fusion of enhanced data sets and global features can greatly improve the stability and generalization of the training model,and is suitable for the capture saliency analysis of a variety of complex models.In addition,as the noise scale of the model increases,the network can still predict a reasonable saliency region,so this method has a certain anti-noise ability.