Saliency Detection Algorithm Based On Multiple Information Aggregation

Saliency detection is a type of image processing task that aims to identify the most obvious objects or regions in an image and segment them accurately and completely.Since saliency detection can help humans or computers quickly capture the most conspicuous or valuable information from images,it is widely used in preprocessing step in visual applications such as visual tracking and image retrieval.Recent years,with the development of deep learning saliency detection based on Convolutional Neural Networks have also achieved significant improvement.However,Convolutional Neural Network has the disadvantage that the shallower layers retain more complete details but lack of semantic information,while the deeper layers contain stronger semantics but sacrifice spatial resolution.This further leads to the problem that it is difficult to obtain accurate saliency maps when only using one level of features.Therefore,whether the multi-level features can be used effectively becomes one of the key factors affecting the detection results.To this end,from the perspective of using multi-level features,this paper first proposes a Stepwise Aggregation network based on the encoder-decoder structure.With the help of the stepwise aggregation module,the network selects different fusion methods according to the type of information,and gradually extracts higher-quality information from the multi-level feature maps and the temporary saliency map that generated in the previous stage to help detection.Through ablation experiments,this paper analyzes the impact of multi-level features and temporary saliency maps on the saliency detection results,and proves that adopting stepwise aggregation method is more effective when integrating multiple information.In addition,this paper also designs a Triangular Aggregation Network with maintaining the advantages of low-level features.The network adopts a brand-new combination of seriesparallel structure and uses the series structure to extract features.Then,the network establishes multiple parallel branches to maintain the underlying information.After building this structure,the performance of the network has achieved a significant improvement.Finally,this paper tests the network on five widely used datasets.And the experimental results show that the triangular aggregation network achieves the state-of-the-art performance.