基于多尺度特征融合与分块注意力的齿轮表面缺陷分割算法

Key Points

• Develop a segmentation algorithm to improve defect detection in gear surfaces amidst complex backgrounds.
• Implemented a multiscale feature enhancement module replacing standard down-sampling in UNet.
• Utilized parallel multi-branch convolution to extract features from multiple scales and directions.
• Introduced a block feature focusing module with multi-head attention for independent local analysis.
• Designed a weighted combined loss function incorporating Dice loss, BCE loss, and gradient difference constraints.
• Achieved an accuracy of 91.27% and 85.88% for various gear defect segmentation tasks.
• Demonstrated superior performance compared to UNet and other advanced models in experiments on a self-built and public defect dataset.