Research On Low Illumination Semantic Segmentation Method Based On Attention Feature Alignment And Domain Adaptation

Semantic segmentation is one of the key tasks in the field of computer vision,which has been widely applied in many areas such as autonomous driving,video surveillance,and medical image processing.Currently,deep learning-based semantic segmentation algorithms for normal illumination have achieved great success,but research on semantic segmentation algorithms for low illumination images is still limited.In low illumination images,the color and brightness of objects significantly decrease,and the image is characterized by low brightness,low contrast,low signal-to-noise ratio,as well as uneven illumination,blur,and loss of detail information.These characteristics of low illumination images make precise annotation of low illumination images extremely time-consuming and laborious,making it even more difficult to create large-scale low illumination semantic segmentation datasets.To address these issues,this thesis studies the low illumination semantic segmentation task from two perspectives,namely,fully supervised and unsupervised domain adaptation,and proposes two different low illumination semantic segmentation methods as follows:To address the issue of poor segmentation performance when using existing semantic segmentation methods directly in low illumination scenarios,we propose a low illumination image semantic segmentation method based on the Attention Feature Alignment Pyramid Network(AFAPNet).First,ConvNeXt is used as the backbone network to extract features from low illumination images.Second,the output feature map of the last stage of ConvNeXt is input into the receptive field enhancement module to capture features of different scales and shapes.Third,the output feature maps of the first three stages of ConvNeXt are input into the multi-branch attention optimization module to capture multi-scale features and reduce the network’s sensitivity to noise in low illumination images by highlighting key information.Finally,the attention feature fusion module efficiently fuses the high-level semantic information of the deep feature maps with the low-level detail information of the shallow feature maps,achieving cross-level information complementarity and obtaining more abundant and discriminative feature information.Experimental results demonstrate that AFAPNet achieves end-to-end low illumination image semantic segmentation and achieves good segmentation accuracy.This thesis proposes a low illumination semantic segmentation framework based on a Low Illuminance Domain Adaptation Pyramid Network(LIDAPNet)to address the issue of scarce labelled samples in low illumination images.Knowledge is transferred from the normal illumination source domain to the low-light target domain by establishing a mapping relationship between normal illumination image samples and low-light image samples.Firstly,pixel-level blending is adopted to mix images from different domains and form new enhanced samples to achieve cross-domain consistency of input image samples.Secondly,ConvNeXt is used as the backbone network to extract features of two illumination domains to achieve domain alignment at the feature level.Then,the output feature map of the last stage of ConvNeXt is sent to the context information aggregation module to enable the network to perform domain adaptation on discriminative regions with contextual information.Finally,the long-distance key information fusion module is used to fuse multi-level semantic feature information and local feature information to capture semantic information and local features in different illumination domains,thereby assisting the network to achieve alignment at the feature level.Experimental results show that LIDAPNet achieves better segmentation performance on unlabeled low illumination datasets.