Research On Semantic Segmentation Method In Low Light Scene

Semantic segmentation technology is an important research content in the field of image processing and machine vision,and plays an irreplaceable role in many important applications,such as automatic intelligent driving,robot scene modeling,and smart healthcare.Semantic segmentation is a pixel-level semantic classification process for input images,and is an excellent method for scene understanding.Different types of segmentation algorithms can be used depending on the usage scenario,and semantic segmentation of urban streetscapes is the most extensive aspect of research.However,the existing street view semantic segmentation only focuses on sunny scenes with sufficient brightness,which will not achieve the desired results if it is applied to low-light scenes.Aiming at the problem,combines the image feature learning and transformation methods and conduct in-depth research,which we propose two deep networks for semantic segmentation of low-light scenes.This paper proposes semantic segmentation algorithm based on image enhancement for direct segmentation to handle the problems of brightness shortage and information loss in lowlight scenes.The method uses a cascaded semantic segmentation network for end-to-end semantic segmentation,with low-light images as the network input and semantic map as the output.The cascaded segmentation network is mainly composed of two parts,an enhanced network and a semantic segmentation network.The enhanced network is a deep network we proposed to improve the brightness of the low-light images.The proposed model combines the low-level and high-level features of the input,which can improve the brightness of the input while maintaining the color without distortion.In order to train the network,this paper proposes a method for generating low-light scene dataset,which can produce semantic segmentation dataset for network training.Finally,through the joint training of the enhanced network and the existing segmentation network,a cascaded segmentation network is generated,which can directly work on low-light inputs and obtain good segmentation results.The feature information learned from the low-light image contains insufficient brightness information and deviated semantic information,in order to tackle the problem,this paper proposes a low-light scene semantic segmentation algorithm based on transfer learning.The model structure of the method is a generative adversarial network,and its purpose is to complete the process of feature transfer of low-light images before semantic segmentation.Feature transfer is to add important semantic feature information to the low-light features,such as brightness and color.The enough accurate semantic information will enhance the segmentation.result.The discriminator of the proposed network could supervise and facilitate the feature transfer part and complete the correct feature conversion process.The experimental results show that the proposed method can make full use of the feature information of the input images to achieve accurate and effective semantic segmentation in low-light scenes.