Salient Object Detection Approach Based On Deep Convolutional Neural Network

The main purpose of salient object detection is to detect and segment the most attractive objects in an image.This part of the salient object contains a lot of the most valuable information in the image.Therefore,only the salient region of the image need to be processed in many computer vision tasks,and then the processing capacity can be greatly reduced while achieving higher task performance.To date,salient object detection has been used as a pre-processing operation for many computer vision tasks,such as object recognition,image caption,and image retrieval.Moreover,the deep convolutional neural network can extract general deep features from the image.These deep features are very useful for visual understanding,so a large number of computer vision tasks begin to be processed using deep convolutional neural networks,including salient object detection.In this paper,we mainly propose two salient object detection algorithms based on deep convolutional network.The first algorithm constructs a two-stream salient object detection network based on multi-stage refinement.In this network,layer-wise recurrent mechanism and channel attention modules are introduced to improve the detection performance.In addition,we also use iterative training methods to learn more complementary features at different stages.In the end,this algorithm has achieved better results than most state-of-the-art methods on several benchmark datasets.The second algorithm constructs a two-branch multi-task learning deep convolutional neural network for salient object detection.This algorithm uses object contour detection task to assist salient object detection task.In this algorithm,we first introduce a residual module based on context contrasted local features to improve the results of salient object detection.Then a feature interaction module is designed to perform multi-task learning and a sparse convolution module is added to the feature interaction module to improve generalization performance.In addition,we used a mechanism of alternating training on each mini-batch of data to enable the two tasks to gradually learn each other's characteristics without considering the trade-off between two tasks' losses.In the end,this algorithm has achieved better results than most state-of-the-art methods on several benchmark datasets.