Research On The Semantic Segmentation Algorithm Of Low-Visibility Images In Foggy Scenes

A Semantic segmentation technology has a wide range of applications,including medical imaging,geological surveying,robotic vision,and autonomous driving.However,most semantic segmentation models are trained on clear and sunny images,leading to a decrease in accuracy when segmenting blurry images.In daily life,adverse weather conditions such as heavy fog are common,which can significantly impact the clarity of images captured by monitoring equipment.Although there have been many breakthroughs in foggy image enhancement algorithms,they have not been extended to semantic segmentation tasks.To further advance autonomous driving technology and intelligent transportation,it is essential to study semantic segmentation methods for foggy scenes.This thesis focuses on the semantic segmentation of low-visibility images in foggy conditions,and the main work includes the following aspects:(1)A Back-projection Multi-scale Feature Fusion Network(BMFF-Net)is proposed for the segmentation task of foggy images.The back-projection multi-scale feature fusion method retains spatial and detail information from different stages of deep networks and strengthens the feature information of targets affected by fog noise through dense feature fusion.The network achieves a mean Intersection over Union(m Io U)of 70.5% on the Foggy Cityscapes dataset and 79.1% on the Cityscapes dataset.(2)An attention mechanism is introduced to improve the segmentation performance of foggy images.Based on the compressed excitation attention module,a parallel branch with maximum global pooling is added.Various pooling methods jointly extract more comprehensive global information,and max-pooling can compensate for the shortcomings of average pooling in noisy images.Using BMFF-Net as the base model,the modified attention module is inserted at different positions,improving the m Io U on Foggy Cityscapes by 1.9% and significantly enhancing the segmentation of small targets.(3)A feature distillation method with attention mechanism is proposed to enhance the network’s segmentation performance for foggy images using defogged features.An enhancement module-based defogging branch structure is added after the network encoder,connected in parallel with the decoder,using the mean squared error loss function as the reconstruction loss.The intermediate features obtained through the attention-based feature distillation method help guide the decoder features,resulting in defogged features more suitable for semantic segmentation.The BMFF-Net with this method achieves a 2.4% increase in m Io U on Foggy Cityscapes,and an ideal score of 49.7% on the real-world foggy image dataset Foggy Driving.(4)Considering the practical application scenarios of foggy image semantic segmentation tasks,lightweight modifications of the network are investigated.Comprehensive analysis of segmentation accuracy,network parameters,FLOPs,and FPS leads to the replacement of the residual group structure in the BMFF-Net with the lightweight feature extraction structure Shuffle Net-v2.This optimized lightweight network for foggy image segmentation has 9.23 MB parameters and achieves a segmentation performance of 63.8% on Foggy Cityscapes and a processing speed of 35.7 FPS on a 2080 TiGPU.