Research On GAN-Based Semantic Segmentation Algorithm For Road Scene Images

In computer vision,semantic segmentation is a crucial technique that helps computers identify objects and their surroundings more accurately in images.In the realm of autonomous driving,it is crucial that driving systems possess the ability to comprehend road scenes down to the smallest details,effectively predicting the class and location of every target in an image through semantic segmentation.To achieve this,it is essential to improve the accuracy and real-time performance of the network as road scenes change rapidly.The focus of our research paper is on semantic segmentation in road scenes.This paper delve into the challenge of balancing network accuracy and complexity,as well as the balance between accuracy and the number of labels required.The approach in this paper is to use deep learning techniques to build semantic segmentation networks with a focus on:(1)To address the problem that existing semantic segmentation algorithms have a large number of model parameters and are difficult to deploy on mobile devices,this paper proposes a lightweight semantic segmentation algorithm.The algorithm improves on Deep Labv3+ by using lightweight Mobile Netv2 for the backbone network to simplify the model and extract low-level feature detail information at different depths;DASPP is used to enhance feature extraction of multi-scale semantic information to obtain rich global contextual semantic information;after introducing an attention mechanism to enhance feature learning,a feature fusion network that preserves detail information is used DFFN was used for feature fusion.In the experiments conducted on the Cityscapes and Camvid,the improved Deep Labv3+algorithm has a model parameter count of 2.74 M and 2.73 M respectively,which is 42.87 M less than Deep Labv3+.The number of model parameters is reduced.This paper presents a lightweight semantic segmentation algorithm model for road images,which can segment images well while significantly reducing the number of model parameters.(2)A semi-supervised semantic segmentation algorithm based on GAN is proposed to address the problem that semantic segmentation algorithms require a large number of finely annotated datasets.Firstly,the discriminative network of GAN is used to generate pseudolabels to assist the generative network for semi-supervised training.Secondly,to improve the training stability of the GAN,a spectral normalization method is introduced into the discriminative network of the GAN.Finally,a modified Deep Labv3+ network is used as the generative network of the GAN,and Res Net101 is used as its backbone network to extract image features initially;MASPP is proposed to enhance feature extraction from a multi-scale perspective for feature maps to obtain rich global semantic information;a feature fusion network DFFN,which preserves detail information,is used for feature fusion to better recover image edges.The Cityscapes and Camvid are used as training and evaluation set for the experiments.On both datasets,the m Io U of the GAN-based semi-supervised semantic segmentation algorithm was 70.9% and 62.0%,respectively,which were 2.5% and 0.8% better than the baseline network,while the algorithm m Io U was 60.1% and 44.7% at 1/8 data volume,respectively that still segmented the images well.The GAN-based image semantic segmentation algorithm proposed in this paper can achieve good segmentation results even under the condition that the number of image labels is small.This study investigates the semantic segmentation algorithm of road scene images based on generative adversarial networks to achieve a semantic segmentation algorithm with lower data integration cost and a smaller number of network model parameters,while ensuring the accuracy of image segmentation.It is of great value and significance to promote the rapid development of autonomous driving by providing accurate road condition information for selfdriving cars while helping to deploy the algorithm in mobile devices.