To address the problem of steel strip surface defect detection, a feature fusion–based preprocessing strategy is proposed based on machine vision technology. This strategy can increase the feature dimension of the image, highlight the pixel features of the image, and improve the recognition accuracy of the convolutional neural network. This method is based on commonly used image feature extraction operators (e.g., Sobel, Laplace, Prewitt, Robert, and local binary pattern) to process the defect image data, extract the edges and texture features of the defect image, and fuse the grayscale image processed by the feature operator with the original grayscale image by using three channels. To consider also computational efficiency and reduce the number of calculation parameters, the three channels are converted into a single channel according to a certain weight ratio. With this strategy, the steel plate surface defect database of NEU is processed, and fusion schemes with different operator combinations and different weight ratios for conversion to the single channel are explored. The test results show that, under the same network framework and with the same computational cost, the fusion scheme of Sobel:image:Laplace and the single-channel conversion weight ratio of 0.2:0.6:0.2 can improve the recognition rate of a previously unprocessed image by 3% and can achieve a final accuracy rate of 99.77%, thereby demonstrating the effectiveness of the proposed strategy.

中文翻译：

---

基于特征融合的钢板表面缺陷识别预处理

为了解决钢带表面缺陷检测的问题，基于机器视觉技术，提出了一种基于特征融合的预处理策略。该策略可以增加图像的特征尺寸，突出显示图像的像素特征，并提高卷积神经网络的识别精度。该方法基于常用的图像特征提取算子（例如，Sobel，Laplace，Prewitt，Robert和局部二进制图案）来处理缺陷图像数据，提取缺陷图像的边缘和纹理特征以及融合灰度图像由特征运算符使用三个通道对原始灰度图像进行处理。为了同时考虑计算效率并减少计算参数的数量，根据一定的重量比，将三个通道转换为单个通道。利用这种策略，对NEU的钢板表面缺陷数据库进行处理，并探索了具有不同算子组合和不同重量比的融合方案，以转换为单通道。测试结果表明，在相同的网络框架下，以相同的计算成本，Sobel：image：Laplace融合方案和单通道转换权重比为0.2：0.6：0.2可以提高以前未处理的识别率。图像的准确率降低了3％，最终的准确率达到了99.77％，从而证明了所提策略的有效性。并探索了具有不同算子组合和不同权重比的融合方案，以转换为单通道。测试结果表明，在相同的网络框架和相同的计算成本下，Sobel：image：Laplace融合方案和单通道转换权重比为0.2：0.6：0.2可以提高以前未处理的识别率。图像的准确率降低了3％，最终的准确率达到了99.77％，从而证明了所提策略的有效性。并探索了具有不同算子组合和不同权重比的融合方案，以转换为单通道。测试结果表明，在相同的网络框架下，以相同的计算成本，Sobel：image：Laplace融合方案和单通道转换权重比为0.2：0.6：0.2可以提高以前未处理的识别率。图像的准确率降低了3％，最终的准确率达到了99.77％，从而证明了所提策略的有效性。