Research On Lightweight Image Semantic Segmentation Algorithm Based On Deep Learning

The main task of image semantic segmentation technology is to distinguish objects of different categories.After setting the corresponding pixels for each category in advance,the input image is predicted through a network,and each pixel is assigned a label.In recent years,image semantic segmentation has gradually been applied to various fields and has high application value in lane line segmentation and medical imaging.However,most existing deep learning-based image semantic segmentation methods cannot meet real-time requirements and it is difficult to achieve a balance between speed and accuracy.This thesis mainly focuses on three aspects of multi-scale information,feature fusion,and network lightweight in image semantic segmentation research.It also analyzes the challenges that exist in practical application scenarios.The main research goal of this thesis is to construct a lightweight image semantic segmentation model by applying relevant content of deep learning after conducting extensive theoretical research.The main work of this thesis includes:1.Sorting out existing real-time semantic segmentation methods and designing a residual bottleneck module which can extract features with fewer parameters when combined with Ghost module,reducing redundant information generated during feature extraction stage thus improving model performance.2.Designing a new dual-branch semantic segmentation algorithm based on an encoder-decoder structure.The high-resolution branch can obtain rough semantic information while preserving detailed information.The low resolution branch line is responsible for generating fuzzy boundaries and obtaining most of the semantic information,which is then fused together through a feature fusion module to generate a feature map with richer semantic information.3.A lightweight semantic segmentation network based on multi-scale feature fusion is designed,which can be divided into two paths: spatial path and global path.The spatial path uses a relatively shallow network to obtain enough spatial information to help restore the high-resolution feature map,while the global path uses a deeper network to extract rich contextual information.The information of the two paths is fused and input into the decoding module,and the bottleneck block unit is improved to improve the overall performance of the model.4.Validating our proposed algorithms effectiveness on Cityscapes dataset by comparing them with existing real-time semantic segmentation networks and analyzing experimental results.Through ablation experiments,the proposed algorithm model has been proven to have good credibility and efficiency.Detailed experiments on AGVS airport dataset also demonstrate the rationality of improved residual bottleneck units.The experimental results show that our proposed lightweight semantic segmentation algorithm can achieve faster speed while maintaining accuracy.