基于多尺度残差网络优化的工业品表面缺陷检测

Surface Defect Detection of Industrial Products Based on Multi-scale Residual Network Optimization

摘要:

工业品表面缺陷检测是工业产品质量评估的关键环节,实现快速、准确、高效的检测对提升工业产能具有重要意义。本文针对传统神经网络提取特征尺度单一、参数量大,网络训练效率低等问题,提出了一种基于残差网络的多尺度特征融合与RBN结合的残差网络模型。首先该模型通过多尺度卷积特征融合模块提取不同尺度的特征信息；然后,通过引入RBN层,使特征分布更加均匀；最后,采用全局平均池化代替传统的全连接层来减少模型的参数量,实现输出通道与特征类别的直接映射。本文提出的网络模型在公开数据集NEU-DET上进行实验,识别率达到100%,在天池人工智能大赛铝型材缺陷数据集上的识别率达到98.8%,模型性能较为优异,可以很好的完成工业品表面缺陷检测任务。

Abstract:

Surface defect detection of industrial products is the key link of product quality assessment in industry. To realize rapid, accurate and efficient detection is of great significance to improve industrial capacity. In this paper, a residual network model based on the combination of multi-scale feature fusion and RBN is proposed to solve the problems of traditional neural networks such as single feature scale extraction, large number of parameters and low network training efficiency. Firstly, the multi-scale convolution feature fusion module is used to extract the feature information of different scales. Then, by introducing RBN layer, the distribution of feature server is more uniform. Finally, global average pooling is used instead of the traditional full connection layer to reduce the parameters of the model, and the direct mapping between output channels and feature categories is realized. The network model proposed in this paper was tested on NEU-DET, an open data set, and the recognition rate reached 100%. The recognition rate on aluminum product defect data set of Tianchi artificial intelligence competition reached 96.67%. The model has excellent performance and can well complete the task of industrial product surface defect detection.