Research On Deeplab V3+ Semantic Segmentation Algorithm Based On Attention Mechanism

In human daily life,image semantic information plays an important role in the transmission of non-verbal information.With the rapid development of Convolutional Neural Network(CNN)and deep learning,research on Semantic Segmentation has achieved fruitful results in the fields of computer vision,medical image segmentation and autonomous driving.However,with the deepening of the convolutional neural network structure,the dependencies between local information and global information,edge information and context information in the image segmentation process are constantly attenuated,and the results of semantic segmentation accuracy are not ideal.Therefore,it is necessary to improve the long-term dependency between semantic segmentation information.The study of relationships is of great significance.With the continuous progress of computer vision technology,the attention mechanism model,which plays a very important role in natural language processing,has also been widely used in semantic segmentation.Self-Attention and Channel Attention find the correlation between semantic information through two different methods,highlighting the important features of the segmented objects.In this way,the long-term dependence of semantic information under the deep network structure is strengthened.Aiming at the above problems,this paper designs semantic segmentation methods of self-attention model and multi-level channel attention model based on attention mechanism.Both semantic segmentation models use an encoding-decoding structure.In the encoder,the high-level semantic features through the deep convolutional network are pooled in the empty space pyramid and then upsampled,and then spliced with the low-level semantic features in the decoder.Upsampled output.For the semantic segmentation method of the self-attention model,we introduce three key values Query,Key,and Value after the high-level semantic features.By calculating the weight coefficients of Query and Key,and then weighting and summing Value according to the weight coefficients,the segmentation model is more efficient.Pay attention to important information.For the multi-level channel attention model,we introduce a channel attention network in the decoder,and recalibrate feature maps through max pooling and global average pooling to obtain cross-channel spatial scale information.In order to verify the performance of the semantic segmentation method based on the selfattention model and the semantic segmentation method based on the multi-level channel attention model,this paper conducts experiments based on the Cityspaces、PASCAL VOC 2012 datasets.The experimental results show that the semantic segmentation method based on self-attention and multilevel channel attention model designed in this paper achieves the mean intersection ratio(m Io U)of72.21% and 73.78% respectively on the Cityspaces dataset,and the improvement effect is obvious.