FAS-YOLO: A Lightweight Network for Steel Defects Detection Based on Deep Learning

Abstract

Timely defect detection is an important means to ensure product quality in the production process of steel products. Detection methods based on computer vision are currently widely popular due to its fast speed and high accuracy. However, computer defect detection methods face challenges of sufficient training within limited samples and balancing detection accuracy and model size. Hence, we propose a lightweight steel surface defects detection model based on YOLOv8, named FAS-YOLO, “FAS” is the abbreviation for our improvement method. Firstly, we introduce a novel convolution method called Falcon to replace the standard convolution of the head component, which can reduce the parameter count and the computational cost of the model. Secondly, we propose the Adaptive Weight Normalized Wasserstein Distance (AW-NWD) loss function. Building upon the Normalized Wasserstein Distance (NWD) loss function, we design a weight coefficients matrix using the Softmax function to establish the relationship between the shape information of the object and the error dimension. Finally, the attention module SimAM is integrated to enhance the model's feature extraction capability without increasing the computational complexity. Experimental results show that compared to the original YOLOv8, FAS-YOLO reduces the parameter count by 20% and the floating point operations by 31%, while achieving an improvement in mean Aver-age Precision (mAP) of 3.2% and 1.2% in NEU-DET and GC10-DET datasets. Furthermore, the proposed model demonstrates superior detection accuracy and reduced computational complexity compared to other YOLO series models, rendering it more suitable for industrial production.