A Visual-Semantic Interactive Carcass Spine Interface Segmentation Method

Fast and efficient segmentation of pork carcasses is of great significance. However, the problem of incomplete feature information occurs in pork carcass segmentation. In order to improve the productivity, the pork carcass segmentation algorithm needs to have good real-time performance. To solve the above problems, this paper proposes a visual-semantic interactive carcass spine interface segmentation method. First, the spine interface features of pork carcasses are extracted using serial superposition dilated convolution. Jump connections are added to avoid gradient disappearance. Secondly, the spatial semantic features are learned by combining the semantic relations between different parts of the pork carcass spine interface. The problem of incomplete information of pork carcass spine interface features is solved. Finally, the visual features and spatial semantic features of pork carcasses are fused to achieve fast segmentation of pork carcasses. The experiments show that the algorithm achieves 70.3% mean intersection over union and 0.217s segmentation time in the case of fuzzy pork carcass spine interface, which is a good balance between segmentation accuracy and speed when compared with existing algorithms.