摘要 在工业生产中产品的质量一直是生产的关键钢材作为我国重要的原材料其表面质量会直接影响到最终的产品质量。最初的钢材表面缺陷检测主要是以人工为主不但对人力的要求高而且很难满足需求后来出现的传统缺陷检测方法主要是针对特定材料的检测无法全面的满足缺陷检测任务。在深度学习兴起后基于深度学习的检测算法具有较高的精度和满足实时检测的速度在工业场景中得到了广泛的应用。然而目前仍存在因工业场景较复杂而导致算法检

Abstract

In industrial production, product quality has always been the key to production. As an important raw material in China, the surface quality of steel directly affects the final product quality. Initially, manual detection was the main method for steel surface defect detection, which not only required high human resources, but also was difficult to meet the needs. Later, traditional defect detection methods were mainly aimed at specific materials and could not fully meet defect detection tasks. With the rise of deep learning, deep learning-based detection algorithms have high accuracy and can meet real-time detection speed, and have been widely used in industrial scenarios. However, there are still problems such as poor detection effects due to complex industrial scenarios, which pose great challenges to the application of target detection algorithms in industrial defect detection. In this paper, based on the YOLOv7-tiny target detection algorithm, we conducted corresponding research on steel surface defect detection tasks. The main work of this paper is as follows: (1) For the problem that the size, scale, and type of targets in steel defect detection lead to inaccurate classification and inaccurate positioning, we studied the BiFPN structure in Efficientdet and designed a module based on weighted multi-scale feature fusion: Bicat module. This module adds different weight values to the feature maps of different scales, making them most suitable for predicting different targets. On this basis, we also used shallower modules for weighted multi-scale fusion, so that the network can learn more accurate location information. (2) In response to the problem of unreasonable allocation of computing resources in the network model, this paper studied the SENet, CBAM, and CA attention mechanisms, and improved the YOLOv7 network model based on this. This strengthens the extraction of target audio information and location information in steel surface defects, and allocates more network model computing resources to defect targets. (3) In response to the limited hardware resources and real-time detection requirements in the current actual industrial production scenarios, this paper compresses the improved model through model quantization and knowledge distillation, achieving a reduction in model size and an improvement in model inference speed while ensuring a certain detection accuracy. To meet the requirements of industrial detection, a model simulation system was designed to achieve the purpose of steel surface defect detection.

Keywords: steel surface defect detection; multi-scale fusion; attention mechanism; YOLOv7-tiny; model compression.