Research On Progressive Salient Object Detection Algorithm Based On Nearest Neighbor Optimization Calculation

Salient object detection aims to automatically localize the most visually conspicuous and distinctive regions that attract human attention in patterns through algorithms.It can be commonly served as a preprocessing step to a variety of computer vision applications including scene classification,visual tracking,video compression and action recognition,to reduce the complexity of subsequent operations,relieve the interference of redundant information and improve the efficiency of image processing.In this thesis,firstly,a structural features combined with short connection network and an independent bidirectional annular feature pyramid network are proposed to enrich the extraction of features,promote the interactive fusion of information and serve as a basic framework for subsequent work.Then,in order to alleviate the problems of missed detection of foreground targets and blurred edge contour,a lateral hierarchically refining network and a progressive two-stage separable architecture are proposed,both of which are based on the nearest neighbor algorithm.The main contribution of this thesis can be concluded as follows:1.In order to enhance the sensitivity to foreground objects,a multi-scale feature extraction module is proposed to extract directional guidance structural information from different receptive fields.It is combined with the skip-type short connection information fusion strategy to accelerate the information flow and promote the mutual complementation of deep-level low-resolution context semantic features and shallow-level high-resolution spatial edge-aware information.In addition,a bidirectional annular feature pyramid network is proposed,which combines top-down and bottom-up annular information flows to enhance the hierarchical representation of features.2.Due to the parametric decision-making mechanism of the convolutional layer and the lack of spatial correlation caused by sliding windows,using a single-scale 11× or 33 × convolution kernel as the final classifier to predict the saliency map will aggravate the problems of missed detection and blurred boundary.The nearest neighbor algorithm is based on the non-parametric decision-making mechanism and global feature contrast calculation.So a top-down and deep-to-shallow multi-level architecture using the nearest neighbor algorithm as an auxiliary optimization process is constructed to gradually restore the internal structure integrity of foreground regions and obtain relatively clear object edge contour,which alleviates the above-mentioned problems to some extent.3.Under the active promotion of the nearest neighbor algorithm,the network architecture is further analyzed and optimized,and then a progressive two-stage separable optimization architecture is proposed,which executes the parallel classifier from multiple receptive fields and the nearest neighbor algorithm in order.It can either use joint supervision to directly generate high-confidence saliency prediction maps end-to-end,or use asynchronous supervision as a post-processing operation to further optimize the speculative results of other algorithms.Qualitative and quantitative experiments were performed on six public available benchmark datasets.Compared with the 18 state-of-the-art deep learning algorithms,the architecture proposed in this thesis has greatly improved the prediction accuracy and generalization capabilities.