Based on DeepLab v3+ model to realize the road scene semantic segmentation

Abstract

The semantic segmentation of deep learning has a very broad development prospect in the field of computer vision, but many network models with good segmentation effect have problems such as large amount of model calculation and long training time for segmentation in road scenes. In response to these problems, this paper changes the feature extraction network to the lightweight MobileNetV2 network, and changes the ordinary convolution in the atrous spatial pyramid pooling module to the depthwise separable convolution, which can effectively reduce the input parameters of the model and greatly reduce the calculation amount of the model. In order to improve the calculation speed of the model without reducing the detection accuracy as much as possible, an attention mechanism is added to the feature extraction module to further optimize the recognition effect of the model for object edges and improve the segmentation accuracy of the model. The experimental results show that the model can better extract the edge features of objects in road scene images and the segmentation accuracy of the optimized model can reach more than 86%.