Research On Salient Object Detection Algorithm Based On Multi-feature Aggregation

With the rapid development of computer vision,image processing technology has been widely used in many fields.Among them,salient object detection is an important research task in the field of computer vision,whose main purpose is to detect the most attractive objects and regions by analyzing the content of an image.In recent years,with the rapid development of deep learning technology,deep convolutional neural network has become the main method in the field of salient object detection.Deep convolutional neural network is used for end-to-end learning and processing of images,thus improving the accuracy and efficiency of salient object detection.However,the current algorithms still have many problems,such as blurry object boundaries,missing small object and feature fusion noise.In view of the above problems,this paper proposes corresponding solutions,and the main research work is as follows:1.Multi-feature aggregation network for salient object detection.To address the problem of object’s blurry boundaries,a salient edge feature extraction module is designed.Considering that shallow networks contain more detailed information,a salient edge feature extraction module is designed in the first layer of the network to obtain clear boundaries.The salient edge feature extraction module not only optimizes the boundaries of salient objects,but also locates salient regions.Due to the convolution characteristics,significant features are diluted in the top-down path.Therefore,a multi-level feature extraction module is designed to preserve the feature information of each layer to improve the integrity of salient objects.The multi-level feature extraction module retains the information of each layer to reduce the loss of small object.Furthermore,a residual refinement module is constructed to further refine the saliency map obtained from the main network by learning the residual between the saliency map and the ground truth.Finally,a hybrid loss function is used to supervise the training of the saliency map produced by the network.Experimental results on four public salient object detection datasets show that the model has demonstrated excellent performance.Especially in the optimization of object boundaries,the model can generate clear salient boundaries.2.Multi-level progressive fusion network for salient object detection.In order to address the issue of excessive background noise generated during the feature fusion process,a novel progressive fusion framework is proposed.Most of the feature pyramid network algorithms suffer from excessive background noise due to a large feature fusion span.To reduce the background noise generated by fusion and enrich the feature information between adjacent layers,a progressive feature aggregation strategy is adopted between adjacent layers.At the same time,a residual channel attention module is designed to assign different channel weights to the aggregated features.To address the losing of small objects during progressive fusion,different levels are connected by skip connections at the same resolution,which reintegrates potentially lost information into the network to improve saliency detection results.Finally,a complementary intersection over union loss function is designed to assign greater weight to foreground objects when saliency objects are small,in order to improve performance of small object detection.The proposed algorithm is evaluated on four public salient object detection datasets,and the experimental results show that the proposed method can effectively improve the performance of detection.