Residual Network Based Salient Object Detection And Its Application

Recently,computer vision has developed rapidly with the unremitting efforts of scholars,and saliency detection has been widely used in machine learning and image processing,including target recognition,image retrieval,image classification,semantic segmentation and visual question answering.However,with the expansion of image datasets and the increase of scene complexity,traditional salient object detection methods have been difficult to meet the needs of researchers,and it is difficult to capture semantic information in many computer vision tasks(such as edge detection and semantic segmentation),which has great limitations.Recently,with the rapid development of deep learning technology,salient object detection has made great progress.However,in the face of complex scenes,the detection effect is still poor,such as low contrast.high similarity of object background,multiple small objects with different characteristics and so on.Therefore,this paper studies a more accurate deep significant object detection network,the main research work is as follows:1.A salient object detection network based on residual feature pyramid is proposed in this paper.Firstly,in order to improve the semantic information of convolutional features,we introduce rich convolutional features to extract more multi-scale and multi-level object information in each side output stage,so as to improve the semantic discrimination in complex scenes.Secondly,the residual feature pyramid structure is proposed to effectively fuse the different side output convolution features based on these optimization.By focusing only on the residual features through the reverse attention mechanism,the missing area and boundary of the object can be effectively learned,and the network redundancy can be effectively reduced and the detection efficiency can be improved.2.In this paper,we propose embedding attention and residual network for salient object detection.First of all,we can learn more accurate semantic features by introducing attention mechanism to guide the feature learning of each side output layer.By learning the attention weight from top to bottom and guiding the shallow object location with high-level semantic features,we can better filter out the interference of background features.Secondly,the residual optimization network is proposed to adapt to the fusion of multi-scale features,and a point multiplication second-order term is introduced to approximate the residual feature in the additive fusion,so as to guide the network to learn the missing salient part of the object.Moreover,such a second-order term can not only facilitate gradient propagation,but also increase network nonlinearity.3.A large number of experimental results show that the method proposed in this paper has more excellent performance compared with many advanced methods in five open data sets and four widely used evaluation indexes.In addition,it has more advantages in operating efficiency.This paper also applies the proposed method to background replacement and cover character replacement,and achieves good experimental results.