Lightweight Semantic Segmentation of Intelligent Workshop Scene Objects Combining Multi-Scale and Attention Mechanisms

Semantic-level segmentation and recognition of scene objects is the basis for realizing intelligent navigation and intelligent security of mobile robots in workshop scenes. Aiming at the multi-scale problems of object semantic segmentation in workshop scenes, such as many types of segmentation objects and large differences in shape, in order to meet the real-time requirements of intelligent workshop object segmentation, an integrating double average pooling and three branch attention mechanism network is proposed. First,an encoder-decoder structure is adopted, and a lightweight convolutional neural network is used as the encoder of the entire network. The decoder includes a two-way average pooling module and a three-branch attention mechanism module to extract the semantic information of multi-scale objects and achieve high-precision semantic segmentation;then, three different lightweight convolutional neural networks,ShuffleNet v2, SqueezeNet, and MobileNet v2, are selected, combined with the decoder, through the object semantic segmentation comparison experiment for the object semantic segmentation dataset of the intelligent workshop scene, MobileNet v2 is determined as the encoder. The experimental results of semantic segmentation accuracy and semantic segmentation real-time performance with lightweight semantic segmentation networks such as ENet, ERFNet, BiSeNet v2, Deeplab v3, Deeplab v3+, CFPNet and Fast-SCNN show that the proposed network achieves 94.25% MPA, 87.67% MIoU, 109×1.66 FLOPs, and 66.67 frame per secondinference speed for object segmentation in workshop scenes, which can well balance the segmentation accuracy and real-time performance, and meet the needs of object semantic segmentation in intelligent workshop scenes.