Saliency Detection Based On Deep Learning In Multi-Data Scenarios

Saliency detection which is called salient object detection.Its purpose is to segment the salient object which is interest to human eyes in videos or images with intelligent calculation and understanding.Because the result of this task simulates the human attention mechanism and can be used as a priori information in many computer vision tasks,this task has received more and more attention from researchers.In recent years,thanks to the application of deep learning technology,saliency detection tasks have developed rapidly.Based on the fully-supervised saliency detection task with RGB image input,the refinement tasks of saliency detection in various scenarios have been proposed one after another.This paper mainly studies the saliency detection algorithm based on deep learning in different data scenarios.For the basic direction of this task and the two refinement directions(RGBD scene and weakly supervised scene),this paper respectively proposes effective deep learning models and frameworks for training and testing.In terms of basic direction,based on the idea of multi-level semantic information fusion,this paper fuses the feature maps in different scales to improve the recognition ability of the network.In addition,this paper aggregates feature maps in the same scale to refine the boundaries of salient objects.Finally,the final saliency result is gradually optimized by grouping recursion.In order to improve the effect of network in complex scenes with spatial information,the paper have conducted research on the method of using the depth information to improve the fully-supervised network results.In a complex scene with deep spatial information,this paper proposes to combine relative depth and absolute depth to help the network slightly improve its effect.The network first incorporates the absolute depth information into the RGB features through the absolute depth module based on the gated unit.After that,the network use the relative distance between the feature points to construct a graph,and introduce the relative depth information into the information flow through the graph network to help the network further refine the saliency results.The above-mentioned saliency detection algorithm in the fully-supervised direction relies on the finely labeled ground truth,and the ground truth map requires a lot of time and labor.At the same time,the difficulty of obtaining the ground truth maps of some scenes also limits the generalization ability of algorithms.In order to reduce the cost of ground truth acquisition,facilitate the training of the network in different scenarios,and improve the generalization ability of the algorithm,The paper research the method of training the saliency detection model using image category labels as supervised information.The paper proposes a contrast-based global pooling method to help the network better obtain the saliency pseudo ground truth through the category label.After that,a network converts the image into a graph with super-pixels as nodes,and propose a spatial graph network to further optimize the pseudo ground truth.Finally,a fully-supervised saliency detection model will be trained using the supervised information with the pseudo ground truth.The above three methods have been tested on multiple datasets.Compared with other algorithms in the corresponding scenario,our algorithms have achieved satisfactory results.At the same time,the ablation experiments in each method also proved the superior performance of our proposed module and further proved the effectiveness of our algorithm.