| Image semantic segmentation,as a fine-grained analysis task in the field of computer vision,continues to make efforts in the transportation technology revolution characterized by artificial intelligence,autonomous driving,and intelligent transportation,and can provide core technical support for environmental perception of traffic scenes.With the continuous progress of economic society and the acceleration of urbanization,problems such as road congestions and frequent accidents are becoming more and more serious.It is an inevitable trend to maximize the efficiency of traffic information resources.Fast and accurate image semantic segmentation technology for traffic scenes is crucial to achieving safe and reliable driving of autonomous vehicles,and timely and effective command and scheduling of intelligent transportation systems.However,practical application scenarios often have complex environments and diverse objects,and visual images often have interference factors such as obscured objects,changing illumination conditions and complicated backgrounds.How to finely segment all objects and object edges,and efficiently complete the segmentation task is the key issue for image semantic segmentation to be applied to actual scenes.Therefore,we focus on the problems of loss of detail information and large number of model parameters in the image semantic segmentation algorithm,and study the image semantic segmentation algorithm based on fully convolutional network and knowledge distillation.The main contents are as follows:(1)Aiming at the problem of low segmentation accuracy of small objects,an improved image semantic segmentation algorithm for small objects based on fully convolutional network and superpixels is proposed.First,the rough semantic segmentation results are obtained based on the fully convolutional network,multi-threshold binarization segmentation algorithm,contour detection algorithm,and internal points extraction algorithm for small-area contours are proposed to obtain the location information of small objects.Secondly,the superpixel segmentation results are obtained based on the simple linear iterative clustering algorithm,the superpixels corresponding to each pixel of the small objects are located,and the category labels of each pixel in the corresponding superpixel are extracted.Finally,the label of the small object is re-labeled as the category with the largest proportion of superpixels.The experimental results on the Cam Vid show that the proposed algorithm can improve the segmentation accuracy of small objects and is effective in improving the performance of image semantic segmentation.(2)Aiming at the problem that the edge of object segmentation is not smooth and small objects are difficult to be accurately segmented,a dual attention guided cross-layer optimized image semantic segmentation algorithm is proposed.First,a coding network with multi-branch feature extraction is constructed.The assignment of dilation rate of atrous convolution is non-proportional and can extract spatial context information,and improve the loss of small object information.Secondly,a cross-layer feature fusion decoding network based on spatial alignment is constructed to fuse semantic information and detail information.Finally,channel and spatial attention mechanisms are proposed to model global channel correlations and long-distance location correlations,thereby enhancing the learning ability of the network for key features.The experimental results on the Cityscapes and Cam Vid show that the proposed algorithm can smooth the edge of object segmentation,especially for the long and thin objects.(3)Aiming at the problems of loss of detail information and large amount of model parameters,an image semantic segmentation algorithm based on edge knowledge distillation is proposed.First,an adaptive multi-scale feature fusion module is constructed,and adaptive weighted fusion method based on each pixel point at different stages is designed to take advantage of each stage of the feature extraction network.Secondly,a spatial attention mechanism embedded in channel attention is constructed to alleviate the conflict between different attention mechanisms.Finally,a knowledge distillation network based on edge information is constructed to improve the edge segmentation accuracy of the lightweight model.The experimental results on the Cityscapes and Cam Vid show that the proposed algorithm can improve the loss of object boundary information,and can achieve a trade-off between segmentation accuracy and model size in image semantic segmentation. |
