Real-Time Steel Surface Defect Detection and Classification with Inference Acceleration

Abstract

Detecting defects on surfaces is a crucial challenge in the steel industry. Various object detection models, including one-stage and two-stage approaches, have been developed to address this problem. Currently, there is a scarcity of models that can effectively and efficiently handle both defect detection and classification tasks in real time. It is widely recognized that striking a balance between inference speed and the accuracy of object detection is a critical aspect that needs to be addressed. To address this challenge, our objective was to develop a model that ensures a high detection rate while achieving real-time processing capabilities. In pursuit of this objective, we conducted a comparative analysis between YOLOv7, a one-stage model, and Faster R-CNN, a two-stage model, followed by model optimization using TensorRT to enhance both inference speed and detection performance. As a result, we have successfully implemented a defect detection model utilizing actual production data, which achieved a detection rate of approximately 98.8% and a false ratio of 20%, while operating at a speed of 46 frames per second (FPS). This achievement demonstrates the effectiveness of our approach in balancing detection accuracy and inference speed.