Abstract: To solve the problems of fuzzy boundary and inaccurate localization of defect features in deep learning algorithms for extracting surface defects of strip steel,a defect detection method called Spatial Multi Scale Weighting (SMA) module was proposed.In this method,the multi-scale feature fusion was utilized to weight different feature information,reducing the weight of background information while alleviating the phenomenon of feature decentralization.In the SMA module,different dilation convolution parallel structures were used to capture features at different scales,while three trainable parameters were used for weighting features at different scales.Embed this module into the tail of the feature extraction network and trained it on the NEU-DET strip steel dataset.The results showed that the backbone network embedded with SMA module had stronger feature extraction ability,and its detection effect was better than the original backbone network.The effectiveness of the proposed method was also demonstrated on another publicly available dataset GC10-DET.

Key words: defect detection, feature fusion, attention mechanism, dilation convolution

摘要： 为了解决深度学习算法在提取带钢表面缺陷时存在缺陷特征边界模糊和定位不准确的问题,提出了一种空间多尺度加权(SMA)模块的缺陷检测方法。该方法利用多尺度特征融合方式,对不同特征信息进行加权处理,减小背景信息所占权重的同时缓解了特征不集中的情况。在SMA模块中,不同的膨胀卷积并行结构用以捕获不同尺度的特征,同时3个可训练参数用于不同尺度特征的加权。将该模块嵌入到特征提取网络的尾部,并在NEU-DET带钢数据集上进行训练。结果表明,嵌有SMA模块的骨干网络具有更强的特征提取能力,相比于原骨干网络,其检测效果更为优秀。同时在另一个公开数据集GC10-DET上也验证了该方法的有效性。

关键词: 缺陷检测, 特征融合, 注意力机制, 膨胀卷积