Real-time and accurate defect segmentation of aluminum strip surface via a lightweight network

On the premise of ensuring the defect segmentation precision of aluminum strip surfaces, improving the segmentation speed to meet the real-time requirements of the production line is an important task. Therefore, a lightweight and efficient network is proposed for the defect segmentation of aluminum strip surfaces. In the network, the lightweight GhostNet with the proposed dilation attention mechanism embedded is used for multi-scale feature extraction. This mechanism focuses more on the critical space and channel features and obtains a large receptive field in the spatial dimension. The Ghost module-based lightweight fusion node is constructed and embedded into the bidirectional feature pyramid network (BiFPN) for more efficient integration of multi-scale features. In addition, a novel lightweight boundary refinement (LBR) block designed as a residual structure is suggested to improve the localization ability near the defect boundaries. The aluminum strip surface dataset with five kinds of common defects is created and adopted to train and test the networks. The evaluation results demonstrate that the mean intersection over union (mIoU) of the proposed network is 85.51%, the speed is 68.86 fps, and the model volume is 9.38 MB. In summary, the proposed network gets a good trade-off between defect segmentation speed and accuracy for aluminum strip surfaces, which provides the potential for real-time segmentation on embedded systems.