Semantic Segmentation of Mechanical Workshop Scenes with Branch-Atrous Convolutional Neural Networks

The semantic segmentation method of mechanical workshop scene is a key technology required for the development of autonomous guided vehicles(AGV)in industrial scenes.Aiming at the problem that AGV needs to accurately identify the passable and impassable area,and the mechanical workshop scene has many and dense target categories,leading to the problem of difficult segmentation,a branch-atrous convolutional neural network model based on the DeepLabv3 deep learning network model architecture is proposed.On the basis of the pre-trained residual networks ResNet-50,the branch structure is first expanded,the atrous convolution with different expansion rates is set through the branch structure to achieve the adjustment of the feature map receptive field and obtain the context information of different receptive fields;then the gridding problem of the atrous convolution is improved through the superposition state of the same expansion rate,the lack of context information is reduced;finally,the model adds a decoder unit of multi-scale feature fusion,the decoder unit uses the shallow features with accurate positioning and the deep features with accurate classification to perform feature fusion,to compensate for the lack of context information and the irrelevance of pixel information due to the gridding problem.The experimental results on the self-made small sample mechanical workshop scene dataset show that compared with the DeepLabv3 model,the verification accuracy of the model is improved by 5.14%,and the results of semantic segmentation for passable areas,road lines and impassable areas are more accurate.