Research On Salient Object Detection Algorithm Based On Multi-Channel Brain-Like Self-Learning

The human visual perception system has the ability to quickly perceive prominent areas or objects of interest in a visual scene.The salient object detection task which is popular in the field of computer vision aims to make the machine simulate the attention mechanism and information screening ability of the human eye,and use limited resources to identify and segment the salient areas or objects of human interest in the image.At present,many depth-based saliency object detection algorithms ignore the difference in the contribution of different features to saliency prediction.It leads to incomplete prediction and background interference in complex scenes.Secondly,the existing salient object detection methods usually only consider the single-channel processing mode,and only use a fixed training dataset to train the model.They rarely consider the self-increment of the training dataset and the self-replacement of the model,and lack the fusion at the decision-making level.This is inconsistent with human visual perception.For solving the above problems,the main contents of this paper are as follows.First,a multi-attention embedded network(MAENet)is proposed for salient object detection.The multi-scale feature extraction module of this network uses 4dilated convolutions with different sampling rates to enrich the multi-scale context information of high-level features,and uses the self-attention module for optimization and screening.And hierarchical multi-attention feature fusion mechanism is used to effectively guide the fusion between deep features and shallow features.Then the attention residual refinement module is designed to optimize the detailed output of features.This paper performs performance evaluation on 5 public datasets,and the results show that MAENet can filter background noise and retains object details better in complex environments,and its performance is better.Second,a salient object detection algorithm based on multi-channel brain-like self-learning is proposed.In order to simulate the human binocular vision system,an algorithm framework with multiple parallel perception pathways and relying on perception positive feedback and visual perception saturation mechanisms to detect salient objects in the image is constructed.Among them,by comparing the object prediction results of different sensing channels,the degree of perceptual saturation of the measured object pixel can be used as the confidence level.The predicted object with high confidence can be used as an automatic annotation of image objects to form a new autonomous training dataset.Using the autonomous training dataset to regenerate a new supervised deep model can amplify the original visual perception channel and continuously improve the system performance until the performance is saturated.In addition,a salient object detection experiment was performed on the COCO dataset without manual annotation,and a new perception model was obtained through the self-annotated training dataset to expand the perception channel in the original system.The results show that the proposed system can effectively improve the performance of salient object detection,and the detection results are more in line with human perception.