Research On Images Semantic Segmentation Method Based On Convolutional Neural Network

Image semantic segmentation is a pixel-level classification task,which plays an important role in many fields such as driverless cars,geographic information systems,and medical image analysis.Traditional segmentation methods are limited by feature extraction methods,and the image segmentation effect is poor in complex scenes.The convolutional neural network achieves good segmentation results in artificially annotated data sets with complex scenes.Convolutional neural network is mainly used for image classification.It can extract semantic information to identify semantic categories,but after many pooling operations,the feature map loses a lot of spatial location information.Semantic segmentation to achieve pixel positioning needs to compensate for the missing spatial information of the high-level feature maps in the down-sampling process.Therefore,how to use the connection between semantic information and spatial location information is an important factor in improving segmentation performance in the process of restoring resolution:(1)The Global Attention Module proposed by the Pyramid Attention Network,which multiplies the sub-high-level features by the channel weights of the high-level features,this method realizes the direct guidance of high-level semantic features to sub-high-level features,completes the transfer of semantic information in features and establishes a connection between high-level features and sub-high-level features.However,Pyramid Attention Network ignores the impact of high-level semantic features on farther low-level features.In response to this problem,this paper proposes a multi-scale global attention up-sampling module,which reconnects multiple global attention module by using multi-scale jump connections.Multi-scale global attention up-sampling module can combine the direct and indirect guidance of high-level features to low-level features to establish a closer connection between semantic information and spatial location information.Combining the multi-scale global attention up-sampling module and the asymmetric convolution module to construct a semantic segmentation method based on the encoder-decoder structure.The asymmetric convolution module is connected to the backbone network to improve the expressive ability of the backbone network.On the decoder side,multiple high-level semantic features use the global attention module to directly or indirectly guide the same low-level feature to generate multiple feature maps.The generated multiple feature maps are fused using two methods: weight-based and cascade-based.In the process of comparing the effect of different weights on the segmentation effect,the comparison of multiple groups of feature fusion methods based on weights proves that indirect guidance is effective in improving the performance of semantic segmentation.(2)In order that the multi-scale global attention up-sampling module is suitable for real-time semantic segmentation,this paper uses ResNet18 and a simplified multi-scale global attention up-sampling module to construct an end-to-end lightweight semantic segmentation network.The core idea is to use the shallow spatial feature extraction module to aggregate the spatial location information of high-resolution features in ResNet.Then,a simplified multi-scale global attention up-sampling module is used to weight high-level semantic features with direct guidance and indirect guidance to spatial location features.In this way,the connection between semantic information and spatial location information is strengthened.The method is simple in structure,has a small number of parameters and only needs a small amount of calculation to achieve real-time semantic segmentation,and has a good segmentation effect.