Image Semantic Segmentation Based On Pyramid Pooling And Attention Mechanism

Image semantic segmentation is one of the main tasks in the field of computer vision.Its purpose is to classify each pixel in the image and predict its corresponding semantic label.With the developing of deep learning research,the semantic segmentation algorithm based on convolutional neural network directly extracts more expressive target feature information from the image through the multi-layer cascaded complex structure,which improves the accuracy of segmentation.In order to extract the feature information of the image more effectively to enhance the segmentation effect,this paper studies the semantic segmentation algorithm based on the deep learning technology of multi-layer feature fusion from the perspective of efficiently using the feature of each level of the convolutional neural network.The main work is as follows:Firstly,a multi-layer feature fusion semantic segmentation method DSPN based on pyramid pooling is proposed.This method uses the high-resolution shallow features extracted by the convolutional neural network to retain more local information and deep features with high semantic information.It is conducive to the characteristics of accurate classification,and effectively solves the problem of loss of segmentation accuracy caused by the loss of detailed information in the process of multiple downsampling of feature maps.The DSPN model uses the lightweight network Mobile Net V2 as the basic network to extract features,and uses the pyramid pooling structure as the last layer of the coding network to extract the global context information of the image.In the decoding stage,the deep features are continuously upsampled and merged layer by layer with the shallow semantic features of the corresponding size,which effectively combines the deep semantic features and shallow local features in the network to improve the accuracy of segmentation.Secondly,this paper further improves DSPN,and proposes a semantic segmentation method A-DSPN based on pyramid pooling and hybrid attention mechanism.This method introduces a hybrid attention mechanism,which combines channel attention and spatial attention to assign weights to different parts of the feature map,enhances the expression of features,improves the global perception of features.At the same time,an auxiliary loss function is added in the training process to optimize the learning process,to ensure the stable convergence of the network and improve the performance of the network.Finally,the method proposed in this paper is verified and evaluated on two public datasets Cam Vid and PASCAL VOC 2012.The experimental results show that the segmentation method proposed in this paper can effectively improve the segmentation performance of the network while reducing the amount of model parameters.