An improved SegNet network model for accurate detection and segmentation of car body welding slags

The prerequisite for realizing robotic self-adaptive grinding of the car body welding slags is to accurately identify the small and randomly distributed welding slags, and to segment their contour from the complex background. In this paper, an improved SegNet network model is proposed to address the challenging problems in small target detection and contour extraction. Both the ability to capture image features and to perceive global and local information is enhanced by adding the context extractor into the SegNet network. Furthermore, the residual structure and BN layer are used to optimize the decoder of the SegNet network, and the Dropout layer is employed to effectively avoid overfitting of the training model. Based on these strategies, the adaptability of the network to multi-scale targets is further enhanced. Experiments on car body welding slags detection indicate that the accuracy rate of the improved SegNet network model can reach 99.5%, which is 2.7% higher than that before improvement. Meanwhile, the boundary pixels at connection area among welding slags, metal plate, and background are accurately segmented, which provides data support for the subsequent robotic grinding of car body welding slags.