A novel vision-based defect detection method for hot-rolled steel strips via multi-branch network

Defects of hot-rolled steel strips will bring a certain effect to the produce appearance and structural strength, even may bring the lose of economy. Accurate defect detection of hot-rolled steel strips could provide an effective support for precise maintenance decision and quality control. However, the defects of hot-rolled steel strips are always against with some challenging factors, such as weak texture, poor contrast, different signal-to-noise (SNR) ratios among different types of defects, etc., which will bring a great effect to accurate pixel-level defect detection. Recently, deep learning has shown a promising performance on image segmentation, especially, U-Net and its variant networks. However, there are also some shortcomings to affect the detection precision, such as the information loss issue caused by multiple pooling operations, insufficient processing of local feature maps, etc. Especially, the information loss issue caused by multiple pooling operations will greatly effect the detection performance on micro defects. To address the above issues, a deep multi-branch defect segmentation network is proposed in this paper for accurate and automatic pixel-level defect location of hot-rolled steel strips. Aimed at the insufficient processing of local feature maps by skip connections, a residual path is proposed to optimize the simple skip connections to acquire the effective context features, and it also could well reduce the effect of semantic gap issue. Meanwhile, to realize feature enhancement of local feature maps, a multi-scale context fusion (MCF) block is proposed to embed into the bottleneck layer to extract multi-scale attention context features. More importantly, to address the information loss issue caused by multiple pooling operations and improve the detection precision on micro defects, a multi-branch network with shared network weights is proposed for accurate multi-scale defect detection. Experiments show that proposed network could acquire a promising segmentation performance compared with other advanced segmentation models.