MSC-DeepLabv3+: A Multi-Scale Contextual Network for Tool Wear Semantic Segmentation

The wear conditions of cutting tools directly affect the processing quality of workpieces and the safety of machine tool operation. Therefore, precise image segmentation of the wear area is the core prerequisite for tool condition monitoring. Aiming at these problems, such as the loss of context information, insufficient spatial feature recognition and high edge missing rate in current tool wear segmentation methods, this paper proposes an improved semantic segmentation model called MSC-DeepLabv3+. This model deeply integrates the Mamba layer with the Atrous Spatial Pyramid Pooling module in the encoder, enhancing the multi-scale global context representation and making up for the shortcomings of traditional convolutional long-range feature association. We design a multi-level shallow feature fusion strategy, which combines the convolutional attention module to adaptively screen and enhance features to suppress the interference of complex textures on the tool surface. Meanwhile, a feature-aware adaptive upsampling module is constructed to dynamically focus on and weight the worn edges, alleviating the problems of high-frequency detail loss and edge blurring. Experiments indicate that the Dice of this model reaches 0.9560, and the segmentation accuracy is significantly better than similar semantic segmentation methods. Meanwhile, under complex cutting conditions, it possesses high precision, which has practical engineering value for the intelligent management of tool conditions in intelligent manufacturing scenarios.