In industrial manufacturing, the precise detection of surface defects in steel is crucial, as it directly affects product quality and production efficiency. Existing methods for detecting surface defects in steel still have many shortcomings. To address this, this paper proposes a novel machine vision model called LCD-YOLO, which leverages deep learning networks and computer vision technologies to provide a new solution for the automatic detection of steel surface defects. Firstly, Linear Deformable Convolution (LDConv) is introduced to solve the dynamic sampling problem, enhancing the model's ability to handle geometric defects with irregular distributions. Secondly, a new C3k2_FC feature extraction network is proposed, which incorporates a designed dual inductive bias coupling mechanism. This innovative mechanism enhances the model's sensitivity to spatial variations by more effectively capturing the local features of steel defects, thereby improving the flexibility and accuracy of feature extraction. Finally, a new multi-branch detection head (Detect_DS) is designed.
We use the NEU-DET steel surface defect dataset to evaluate the effectiveness of LCD-YOLO. This dataset was originally proposed by the research team led by Song and Yan (2013) from Northeastern University [38]. It is commonly used for steel surface defect detection and classification tasks. The dataset contains images of six different types of steel surface defects: scratches. Each defect type consists of 300 images, with each category containing 300 grayscale images. Each image is annotated with the defect type and its location. There are a total of 1800 images, which will be allocated to the training set, validation set and test set in the ratio of 8:1:1. All images are standardized to a size of 200 × 200 pixels. The dataset also includes annotations indicating the defect category and location for each image. Various steel surface defects from NEU-DET are shown in Figure 6.
This paper proposes a novel detection model based on YOLOv11 for automatic steel surface defect detection. The deep learning model addresses challenges in existing steel surface defect detection methods, such as large defect span, poor detection performance for small defects and low accuracy. Experimental results show that compared with the latest YOLOv11 model, the mean average precision value on the NEU-DET dataset is improved by 2.1%, GFLOPs are reduced by 14.29% and the number of parameters is decreased by 6.43%. Heatmap experiments on different steel defect datasets further confirm that the model can accurately capture diverse defect features, while generalization experiments on the VOC2007 dataset validate its effectiveness in broader object detection scenarios. Various experiments demonstrate that the LCD-YOLO model outperforms current mainstream YOLO models in terms of detection accuracy, speed, generalization ability and parameter count.
(1) New C3k2_FC network module: The newly designed dual inductive bias coupling mechanism. This innovative mechanism enhances the model's sensitivity to spatial variations by more effectively capturing the local features of steel defects, thereby improving the flexibility and accuracy of feature extraction. Its dynamic bias adjustment capability enables the model to accurately identify defects even under complex conditions. (2) LDConv: To dynamically adjust the sampling locations for steel defects and enhance the model's ability to handle geometrically irregular defects, LDConv is introduced. (3) Multi-branch shared detection head (Detect_DS): To improve the efficiency and speed of the detection head, the Detect_DS detection head is designed. Different from traditional single-detection-head methods, through the unique hierarchical multi-branch design and shared parameter configuration, the number of model parameters is significantly reduced and the accuracy of the model is improved.
