High Performance Network for Detection of Surface Defects on Hot-Rolled Steel Strips Based on an Optimized Yolo V3

In the hot rolled steel industry, surface defects are a major problem and various hot rolled steel manufacturers are struggling to find cost effective and efficient ways to solve this problem. However, low accuracy, minimal ease of automation, slow detection speed, and limited robustness are still serious problems. For this purpose, this paper proposes a YOLO v3-based algorithm for detecting surface defects on hot-rolled steel strips. One, 4 residual layers (res_units) are formed by applying neural convolutions, batch normalization (BN) and leaky RELU activation to eliminate gradient vanishing and maximize feature extraction and learning ability of the network. Second, a 4-scale feature map operation is used to improve the detection operation, increasing the grid size of the image features obtained from the previous operation to improve detection. The 4 scale detection is used in this study for accurate and stable detection of the surface defects as they are not fixed size targets. Finally, the 3-scale upsampling layer is used, where the spatial resolution of the image is increased before scaling, followed by a concatenation operation that concatenates outputs of the current layer with the previous layer. Both quantitative and qualitative experimental results show that the optimized YOLO v3 model proposed in this paper in terms of mean average precision (mAP) in detecting flaws on hot rolled steel surfaces with a mAP of 100% and a recall of 95% performs better more accurately than the original YOLO v3 network, Faster RCNN and Xception network. The experimental results of the YOLO v3 model proposed in this paper show an optimal and balanced performance, suitable for use as a qualified network that meets the industrial needs of the hot rolled steel strip industry.