基于通道混洗和注意力机制的轻量缺陷检测算法

A Lightweight Defect Detection Algorithm Based on Channel Shuffling and Attention Mechanism

摘要: 缺陷检测在制造业中扮演着确保产品质量和效率的关键角色。由于缺陷通常没有固定形状，并且受光照影响较大，因此基于图像的高精度缺陷检测成为一项极具挑战性的任务。本文针对缺陷图像的特点，提出了一种结合了通道混洗和注意力机制的U型结构的卷积神经网络。首先在跳跃连接中加入注意力门机制来提高网络的泛化能力；其次，针对上采样中容易产生边缘轮廓失真的问题，采用了混合损失计算；最后，因使用了效率更高的编码解码器，在模型参数和浮点运算数较低的情况下，所提出的模型MIoU和F1指标达到了91.89，94.67和64.82，72.93。与FCN、U-Net、U-Net++进行了比较，结果表明所提出的方法在表面缺陷检测领域优于相关方法。

Abstract: Defect detection plays a crucial role in ensuring product quality and efficiency in manufacturing. Due to the irregular shapes of defects and their susceptibility to lighting variations, image-based high-precision defect detection becomes a highly challenging task. In response to the characteristics of defect images, this paper proposes a U-shaped convolutional neural network with a hybrid loss calculation and attention mechanism. Firstly, an attention gate mechanism is introduced in the skip connections to enhance the network’s generalization ability. Secondly, to address the issue of edge contour distortion during up-sampling, a hybrid loss calculation is employed. Lastly, leveraging more efficient encoding and decoding layers, the proposed model achieves MIoU and F1 scores of 91.89, 94.67 and 64.82, 72.93, respectively, with lower model parameters and floating-point opera-tions. Comparative analysis against FCN, U-Net, and U-Net++ demonstrates the superior perfor-mance of the proposed method in the field of surface defect detection.