Research On Image Salient Object Detection Based On Multi-Feature Fusion

As a key preprocessing step for computer vision tasks,saliency object detection is widely used in object tracking,behavior recognition,and image retrieval.For massive picture data,saliency object detection can quickly and accurately locate the content of people’s interest by imitating visual attention mechanism,that is,detecting the most attractive objects in different scenes,therefore it has very important academic research and practical application value.However,traditional saliency object detection algorithms highly rely on manual features,such as texture and color to obtain local details and global information of images,and the generalization and robustness of the detection effect are weak in the face of complex practical application scenes.In recent years,with the continuous development of deep learning and computer computing power,convolutional neural networks have achieved many successful applications for salient object detection tasks.Its accuracy shows obvious advantages over traditional methods,but there are still many unsolved problems,such as with the gradual saturation of deep learning network models in terms of performance,the focus of saliency object detection tasks gradually changes from locating salient object regions to more refined object contour boundaries,which cannot effectively learn to use the multi-level contour features of salient objects;In the face of multi-object detection objects in complex realistic scenes,it is difficult to effectively use the channel and spatial attention mechanism to integrate the global deep feature information to solve the dependence between independent significant objects.To solve the above problems,this thesis uses the multi-feature fusion mechanism to improve the effect of saliency object detection algorithm,and takes into account the requirements of actual detection,and establishes a suitable,operational and verifiable saliency object detection model.The main research contents are as follows:(1)To address the problem that contour information cannot be used effectively,this thesis proposes a saliency object detection method based on the fusion of salient features and contour features.This proposed algorithm is based on the codec structure of deep residual network.In the bottom-up path,it first introduces a fusion method to exchange the information between contour features and salient features,and then adopts recursive structure to enhance fusion for optimization,so as to improve the utilization rate of contour information.Finally,the feature extraction module is used to extract the most valuable upper features from the phased network model,and the truth graph is used for supervised learning to generate the optimal boundary prediction.Experimental results show that the proposed recursive fusion model can effectively fuse contour features and improve the robustness of saliency object detection in complex scenes.(2)To address the problem of poor detection in multiple salient object scenes,a salient object detection method based on global feature and local feature fusion is proposed.The proposed algorithm is composed of multi-class feature capture mechanism and channel space perception network.Firstly,the multi-class feature capture mechanism sets different feature generation modules for the global information in the image to extract the global features,local fine-grained features and edge features in the image.Then,from the perspective of channel and spatial attention mechanism,the channel spatial perception network reasonably fuses multi-class features multiple times to generate salient object detection results.Experimental results show that the proposed algorithm constructs the fusion of multi-class features through the channel space joint attention mechanism,which can effectively enhance the correlation between multiple salient objects and show superior performance in multi-target complex scenes.(3)Design and implementation of image saliency object detection system.Building upon the research in image saliency object detection algorithms,this thesis designs and implements an image saliency object detection system using MATLAB.The system includes some modules,such as algorithm training,performance testing,and evaluation.It contains several representative algorithms,commonly used saliency object detection datasets,and performance evaluation metrics in the field of image saliency object detection.Furthermore,the feasibility and practicality of the system are validated through comprehensive functional testing,demonstrating its usability,visualizability,and installability.The system exhibits practical value for real-world applications.