Research On Salient Object Detection Algorithm For Cross-modal Collaborative Compensation

In the fast-paced and efficient real life,human beings can quickly focus on the salient areas of interest by using visual attention.In order to deal with the impact of massive image data in the era of big data,human beings can extract the salient regional information of images in large quantities by using the computer to simulate the visual attention mechanism of human eyes.Therefore,Salient Object Detection(SOD)has become one of the hotspots in the research field of computer vision.As human eyes like positioning detection technology,SOD algorithm accurately extracts the key and salient areas contained in the image,which lays a solid foundation for more advanced computer vision tasks such as Object Segmentation,Object Classification and Object Recognition.In recent years,the research on SOD algorithm based on RGB modal extracts salient objects through attributes such as color contrast and texture change,but there is still the phenomenon of fuzzy edge information in the division of foreground and background areas.In order to overcome this challenge,edge information is introduced as an important cooperative compensation modal,so as to improve SOD algorithm,increase the accuracy of the model and refine the edge of salient objects.However,in the face of challenges such as relatively complex scenes and illumination changes,the detection effect of SOD model based on RGB modal is not ideal and its robustness is poor.Tracing back to its source,human vision can also perceive the depth distance of the visual scene.The difference of depth distance can make the salient object highlight from the challenges such as complex background or illumination changes.Therefore,the SOD model can be further enhanced through the cooperative compensation of depth modal information.Aiming at different modal information(Edge,Depth),this paper studies the SOD algorithm of cross-modal cooperative compensation.The specific work is as follows:(1)In view of the unsatisfactory effect of the existing SOD algorithm on the edge perception of salient objects,and considering that the high-level features contain advanced semantic information and the low-level features contain rich texture information in the deep neural network,this paper designs a deep network composed of stacked edge-aware modules,that is,the SOD network based on edge modal cooperative compensation.Firstly,Res2 Ne Xt-50 is used as the backbone network to transform the image into multi-object and multi-scale initial salient features.Then,the edge-aware modules are stacked sequentially to fuse the highlevel and low-level information of the image in an asymmetric way,so as to highlight the salient object position.Among them,the network leads out the branch forcing model to learn the edge modal features and performs cooperative compensation in the edge-aware module.Finally,SOD results are outputted by the network.Experiments show that this algorithm can enrich the edge details of salient objects and further accurately locate the salient object areas due to the cooperative compensation of edge modal.(2)In order to overcome the challenges of complex scenes and illumination changes,SOD network based on depth modal cooperative compensation is proposed in this paper.Firstly,Res2 Ne Xt-50 is also used as the backbone network to transform the input image into rich salient features.Then,considering the gradient and smoothness of the saliency of each region in the image,the Soft Attention Module is designed to optimize the rough features extracted by the backbone network.At the same time,in order to deal with the multi-scale problem of salient objects,the Multi-scale Texture Aware Module is proposed to further improve the feature quality.Secondly,the Depth Cooperative Compensator is constructed,which uses the depth modal features of forced network learning to assist in optimizing the salient features in the way of cooperative compensation.Finally,the salient features are fused through the Pyramid Aggregation Module to predict the final SOD result.In addition,this paper also uses this network to demonstrate the necessity of introducing depth modal in the research direction of SOD based on RGB modal,and in order to expand into a general network framework,this paper demonstrates the universality of cross-modal cooperative compensation in the research direction of SOD based on RGB-T modal.Experiments show the effectiveness of the proposed algorithm in three different research directions of salient objection detection.