In order to better analyze the workshop environment and solve the problems of low semantic segmentation accuracy caused by few sample samples,many categories and large scale changes of the target pixel-level segmentation task of the workshop,an improved U-Net workshop scene segmentation model was designed.The improved model used Rep-VGG lightweight backbone network,and introduced the pyramid splitting attention mechanism in the sampling stage of U-net to increase the feature representation ability and reasoning speed of the model.In model training,Dice-Cross Entropy combination loss function was adopted to solve the problem of difficult training caused by unbalanced samples.Experimental data showed that the model could achieve fast,lightweight and high-precision segmentation on the self-built small-sample workshop dataset.
关键词
语义分割/轻量级神经网络/注意力机制/组合损失函数/深度学习
Key words
semantic segmentation/lightweight neural network/attention mechanism/combined loss function/deep learning
维普
万方数据