A Deeplab-Based Segmentation Network for Screw Images

Aiming at the problem that the screwdriver cannot be precisely embedded in the screw groove area during automatic screw removal, we propose an image semantic segmentation model fused with a lightweight convolutional neural network. Based on the classic DeeplabV3 model, a lightweight MobileNetV2 structure is used to replace original feature extractor, and its unique spatial pyramid structure is used for multi-scale fusion of the convolution feature of screw head image, and adding a dual attention module to extract the high-dimensional feature to reduce the loss of detail in segmentation. Finally, deconvolution is used to restore the resolution through the improved decoding network. By comparing our method with the state-of-the-art semantic segmentation network, It turns out that our method has better segmentation performance, with the mIoU up to 94.6%, and the testing time of a picture is 0.12ms, which can meet the demand of real-time task.