Research On Salient Object Detection Based On Deep Learning

Artificial intelligence aims to uses computers to simulate a range of human cognition.In recent years,the convolutional neural network has been successfully applied in computer vision,especially in the fields of object detection,image segmentation,and tracking.Different from object detection,salient object detection aims to attain the most significant object.The research based on deep learning extends the saliency detection from mining the contrast relationship between images to the extraction and application of low-level features,mid-level features,and high-level features.With the help of the powerful feature extraction capability of convolutional neural networks,profound research has been made in the salient object detection algorithm.However,in the existing models,feature extraction almost takes all features in the scene into consideration,which has a great influence on the performance improvement of detection.How to construct effective strategies to extract the effective information from the image and keep the target features completely has become a new research topic.Based on the research of multi-scale fusion network using deep learning,this paper introduces the attention guidance module to design a new multi-scale fusion network.The main work is as follows:Firstly,this paper summarizes the research background and significance of salient object detection,makes a proper comb of the domestic and foreign development status,and leads to the starting point of this research work..Secondly,this paper summarizes the background of saliency detection and deep learning,and the deep fusion network is introduced as an individual important chapter.Through combing and summarizing the previous research work,it laid a foundation for the follow-up work of this paper.Then,based on the deep fusion network,this paper combines the morphological operation to make some improvement,and proposes the multi-scale fusion and non-local attention mechanism based salient object detection(FNAsNet)and the multi-scale inception network(MSINet)respectively.FNAsNet utilizes the multi-scale fusion strategy at the highlevel to add redundant information to ensure the integrity of the target.The non-local attention module filters the features in two dimensions of space and channel,so as to get more effective features.MSINet removes the redundancy along the object edge in the feature map of each layer,and the multi-branch inception module can fuse the filtered features.The experimental results show that the two algorithms have certain advantages for the detection of saliency in complex scenes.All the algorithms mentioned in this paper have been tested on six public datasets,and the experimental results have shown that the proposed methods have certain performance.