Research On Real-time Semantic Segmentation Algorithm Based On Attention Mechanism

In recent years,semantic segmentation algorithms based on deep learning have become the mainstream approach due to their advantages in segmentation precision.The improvement in precision is attributed to the complex model structure of deep neural networks that enable them to obtain larger receptive fields.The combination with the attention mechanism also allows the model to capture the long-range dependencies between pixels and between channels,thus enhancing the feature representation.However,the large effort of computation required by attention modules leads to a decrease in the inference speed of models,making neural networks unable to meet realtime requirements.For the task where images have obvious structured features,some studies have tried to design real-time semantic segmentation networks using a dual-branch downsampling structure to improve the inference speed of models.However,most of these networks cannot provide sufficient segmentation precision because the feature fusion module is too simple.In view of the above problems,the main research contents and innovation points of this thesis are as follows:(1)A Dual Pooling-aggregated Attention Network is proposed,which aggregates information by using average pooling and max pooling on the non-target dimension of features.A Pooling-aggregated Position Attention Module and a Pooling-aggregated Channel Attention Module are designed to effectively reduce the size of features involved in attention modeling.The computational effort required for attention modeling is significantly reduced,and the inference speed of the model is improved.The experimental results show that the Dual Pooling-aggregated Attention Network achieves 91.08% and 85.39% of Mean Io U on Deep Fish and SUIM datasets,and its inference speed is improved by 31.61% compared with similar models.(2)A Swap Attention Bilateral Segmentation Network model is proposed,and the attention mechanism is applied to design the feature fusion module.The spatial detail branch and semantic branch of the dual-branch down-sampling structure are used to extract features respectively,and a Swap Attention Module is designed to extend the receptive fields of the spatial detail branch and the semantic branch to each other’s global contexts.As a result,the features of these two branches are fused,which effectively improves the segmentation precision of the lightweight model.The experimental results show that the Swap Attention Bilateral Segmentation Network achieves 93.65% Mean Io U on the USVInland dataset while maintaining the same level of inference speed as similar models,improving the segmentation precision.(3)Based on(1)and(2),functions of image segmentation in an aquatic autonomous navigation monitoring system are designed and implemented.For the underwater scene segmentation problem,the Dual Pooling-aggregated Attention Network is used to develop the solution.And for the problem of recognizing navigable water surface with structured image features,the solution is developed using the Swap Attention Bilateral Segmentation Network.