Research On Image Semantic Segmentation Algorithm Based On Deep Learning

With the rapid development and popularization of deep learning and artificial intelligence technologies,semantic segmentation as a high-level task in image processing,aims to assign a semantic label to each pixel in the image to achieve a complete scene understanding.It is widely used in fields such as autonomous driving,medical diagnosis,salient object detection,and aerial image analysis.Aiming at two application scenarios of road information extraction and road crack segmentation from remote sensing images,this paper solves the problems of low precision and low efficiency of the model segmentation by expanding the high-level receptive field and combining with multi-scale context features,and improves the segmentation ability of the model to image detail features.The main research work of this paper is as follows:(1)Aiming at the problem that the foreground of the remote sensing image accounts for a small proportion of the background and the occlusion is serious,a remote sensing image road segmentation network combining intensive attention and parallel upsampling is proposed.The network has designed a dense hollow spatial pyramid attention module in the middle part of the encoder and decoder.This module contains two branches of spatial attention and channel attention.The spatial attention branch is used to increase the network's receptive field and obtain dense multi-scale context information.The channel attention branch is used to capture channel information and realize the adaptive selection of channel features.Finally,the two branches are merged to extract richer global context information,and feature screening is performed through the attention mechanism to eliminate Irrelevant information.Considering that the network is likely to cause the loss of advanced feature information during the upsampling process,this model designs a multi-channel parallel up-sampling module in the decoder module.The module combines features of different levels after upsampling,which enhances the ability of the model to combine multi-level information.The recall rate,accuracy rate,precision rate and F1-score of this method tested on the Deep Globe dataset are 80.5%,99.4%,82.1%,and 80.3% respectively.All indicators are better than the current mainstream algorithms,which further improves the segmentation.accuracy.(2)Aiming at the problem that the road crack features are complex,small and easily affected by external conditions,a road crack segmentation model based on multi-level context feature fusion is proposed.The model adds a position attention module behind each layer of the encoder to extract all position information related to the crack feature,so as to enhance the feature of the query target position and improve the model's ability to combine contextual information.At the high level of the network,the global pyramid residual module is used to expand the network's receptive field by using dilated convolutions with different dilation rates,and the global pooling operation is used to extract the global context information of the network,which is conducive to the segmentation of small crack images.In addition,the model uses residual connections which help improve the performance and robustness of the network.Experimental results show that the comprehensive index F1-score of this model on the CRACK500 test set is 1.06% higher than other algorithms,and the F1-score tested on the CFD data set is 2.72% higher than other algorithms,indicating that the algorithm has a certain generalization performance.(3)Aiming at the imbalance between the segmentation accuracy of the road crack segmentation model and the amount of parameters in Chapter 4,a road crack segmentation model based on improved Link Net is proposed.Taking into account the characteristics of Link Net,such as high segmentation accuracy,less occupied parameters and easy to add,this paper improves Link Net and adds a joint guided feature extraction module on this basis to obtain global and local spatial information,through joint The feature extractor performs a series of feature integration,and finally weights the features to the original feature map to achieve feature screening.At the same time,a hybrid dilated convolution module is added to the high-level of the model to expand the high-level receptive field and improve the model's ability to segment small features.The experimental results show that the comprehensive index F1-score tested on the CRACK500 dataset is 0.35% higher than the algorithm model MCFFNet proposed in Chapter 4.On this basis,it reduces the number of parameters,shortens the image processing time,and the segmentation performance is better than other algorithms.The F1-score tested on the CFD data set is 0.82% higher than the highest value,further verifying the effectiveness of the model.