Research On Lightweight Image Semantic Segmentation Technology For Urban Road Scenes

Semantic segmentation is the process of dividing an image into multiple semantic regions.It is necessary to assign clear semantic labels to pixels in each region to achieve object recognition and visual understanding.Semantic segmentation is a dense predictive task.In this process,it is susceptible to interference from factors such as illumination and sample distribution.In the urban road scene,the environment is complex and changeable.The existing semantic segmentation algorithm network model is complex,the calculation is large,the reasoning speed is not high,and the reasoning speed in the landing application cannot meet the real-time requirements.Aiming at the imbalance between segmentation accuracy and reasoning speed in semantic segmentation,this thesis conducts in-depth research and proposes two lightweight semantic segmentation network models.The specific contents and innovations are as follows:(1)From the perspective of model lightweight,this thesis proposes a real-time semantic segmentation algorithm based on encoder-decoder structure.For the encoder,this thesis proposes a non-bottleneck residual unit combining dilated convolution and depthwise separable convolution.The residual unit combines dilated convolution with different dilation rates to extract multi-scale features.For the decoder,bilinear interpolation is used to restore the image resolution.Compared with the benchmark algorithm ICNet,the inference speed is doubled,the number of parameters is reduced by 34 times,and the segmentation accuracy is almost lossless.The algorithm achieves a balance between segmentation accuracy and efficiency.(2)Combined with the task characteristics of semantic segmentation dense prediction,this thesis proposes a real-time semantic segmentation task algorithm,that is,a bilateral real-time semantic segmentation algorithm based on non-local attention mechanism.The algorithm fully integrates semantic information and spatial information by sharing shallow information and feature guidance.In this algorithm,a lightweight asymmetric bottleneck residual unit is designed by convolution decomposition technology,and an efficient encoder is constructed by the LABR unit.Aiming at the problem that the non-local attention mechanism has a large amount of similarity calculation between pixels,a new lightweight attention mechanism is proposed.Compared with the benchmark algorithm ICNet,the reasoning speed of the proposed algorithm is improved by 2.7 times,the number of parameters is reduced by 14 times,and the accuracy is improved by 0.7%.The algorithm achieves an average intersection-union ratio of 70.2%segmentation accuracy on the Cityscapes dataset and becomes a feasible method for performing efficient image semantic segmentation under limited computing resources.