Research On Image Salient Object Detection Based On Multi-source Context And Complementary Aggregation

Salient object detection(SOD)aims to accurately locate and completely segment the most attractive object or region through analyzing the content in the input image.With the ability to rapidly process image data,SOD as a common preprocessing stage and has been widely used in many computer vision tasks to improve the processing efficiency.Recently,convolutional neural networks(CNNs)have been widely used in salient object detection because of their ability to efficiently extract multi-level features.Salient object detection methods based on full convolutional networks have made great progress through aggregating a large amount of multi-scale and multi-receptive field context information to optimize initial multi-level features.However,these methods have certain limitations in the design of multi-scale context information extraction module,multi-level feature interaction manner and attention mechanism,so that their generalization performance is poor in the face of real-world scenarios,which cannot meet the future industrial application.To be specific,the current convolutional neural networks-based salient object detection methods have major shortcomings: 1)The expression ability of context information obtained independently by dilated convolution is limited,leading to the problem of local information loss and reducing the model prediction accuracy;2)Different aggregation manners were not considered in the multi-level feature fusion stage,resulting in insufficient use of complementary information and poor model generalization performance;3)The attention mechanism has a single field receptive and complex design,resulting in high computational complexity of the model and slow reasoning speed.The above problems hinder the development of salient object detection.To address above problems,this thesis focuses on image salient object detection task,starting from the context information,efficient aggregation strategy and attention mechanism,and puts forward corresponding improvement schemes from three different aspects.The main contributions of this thesis are as follows:1.It is proposed to use asymmetric convolutions,complementary dilated convolutions and diamond hierarchical structure to solve the problem of local information loss caused by the introduction of dilated convolution filling rates when context information is extracted based on dilated convolution SOD methods.In addition,the context information extraction module strengthens its relevance by introducing diverse connections between itself.Extensive experimental results show that the extracted strategy can effectively capture high-quality context information to help understand the image content and locate the salient object from the image,which has obvious performance advantage compared with the existing SOD methods.2.It is proposed to use a variety of aggregation strategy to promote the interactive aggregation of complementary information in the multi-level features,and to integrate the multi-level features at the cost of few parameters and computation to generate more expressive feature representations to improve the performance of the model.Extensive experimental results demonstrate that our method has superior advantages compared to existing SOD methods under different evaluation metrics.3.It is proposed to use multi-field receptive attention to select potentially meaningful information in channels from different persepectives.In addition,a hybrid multi-source attention mechanism is developed to improve the ability of the model to collect global information by using the feature diversity strategy,and the attention diagram is obtained by a lighter and more efficient one-dimensional convolution operation.Extensive experiments on five public datasets demonstrate that the proposed method achieves superior performance and has significant advantages in the memory size of the model against other state-of-the-art SOD methods.Figure [32] table [16] reference [85]...