【摘要】 为了满足热轧带钢表面缺陷检测高准确性的要求，以裂纹、点蚀等六种表面缺陷为研究对象，提出一种基于改进YOLOv8的热轧带钢表面缺陷检测算法RCS-YOLOv8。首先，引入改进的感受野与坐标注意力机制，构建RCC(Receptive field & coordinate attention convolution, RCC)模块，以提升密集目标的检测能力。同时融合RCC与C2f结构的CFC(Receptive field & coordinate attention cross-stage fusion, CFC)模块，实现自适应感受野特征提取。针对微小目标，提出浅层多尺度检测模块SBA(Spatial Bi-directional attention, SBA)增强空间特征感知。采用RCS-YOLOv8网络对热轧带钢表面缺陷数据集进行训练与测试，并与原始模型进行了对比分析。实验结果表明，改进后的算法在NEU-DET数据集上的平均检测均值mAP达到了79.8%,F1检测值达到了78%，分别提高了3%与5%，改进算法的浮点运算量减少了2.4%。在GC10-DET数据集上，mAP提高了4.2%。RCS-YOLOv8在提升检测精度的同时保持较快的检测速度，满足工业带钢表面缺陷检测的准确性和实时性的要求。更多还原

【Abstract】 Currently, numerous challenges are faced in the detection of surface defects on hot-rolled strip steel. These include the difficulty of feature extraction caused by the complex background of crazing(Cr) defects, a weak response exhibited by small target defects such as pitted-surface(Ps), a lack of cross-scale fusion resulting from the different scales and morphologies of rolled-in-scale(Rs) defects, and the issues of a small number of samples and an uneven distribution of categories encountered. To meet the requirements for high accuracy of the surface defect detection algorithm for hot-rolled strip, a hot-rolled strip surface defect detection algorithm named RCS-YOLOv8, which is based on an improved YOLOv8, was proposed. Six surface defects, such as crazing and pitting-surface, were taken as the research objects to enhance the detection precision and robustness of surface defects in hot-rolled strip steel. To address the limitations of the original YOLOv8 in small object detection and multi-scale feature representation, improvements were made in three key areas. First, an enhanced receptive field and coordinate attention mechanism were introduced to construct the RCC(Receptive field & coordinate attention convolution) module. Feature extraction is enhanced by improved receptive field coverage and directional sensitivity; thereby, the recognition accuracy for densely packed objects is boosted. The constructed RCC module not only inherits the directional sensitivity of CA but also is enabled to extract detailed information with greater precision through enhanced receptive field adaptability provided by RFAConv(Receptive field attention convolution). Object localization capabilities are enhanced by the CA mechanism through the modeling of long-range dependencies in both horizontal and vertical directions. However, a fixed local receptive field is maintained by CA, which fails to fully optimize the flexibility of feature extraction. To address these limitations, an adaptive approach is employed by the RFAConv module to optimize feature extraction. Compared to traditional attention mechanisms, weights for different positions are learned by RFAConv to adaptively adjust the receptive field size, which effectively enhances local feature representation. Furthermore, the degree of attention paid to features across different receptive fields can be adaptively adjusted by RFAConv. Simultaneously, the CFC(Receptive field & coordinate attention cross-stage fusion) module, which integrates both RCC and C2f(Cross-stage partial fusion) architectures, enhances multi-scale feature representation capabilities while computational efficiency is maintained, enabling adaptive receptive field feature extraction. The core of the CFC module is considered to lie in the refinement of the multiple Bottleneck structural submodules that compose the C2f module. As the fundamental building block of the C2f module, structural refinement is undergone by the Bottleneck structure, where its standard convolutions are replaced with RCC modules. This modification, performed without altering the overall C2f framework, enables adaptive receptive field extraction capabilities; thereby, the feature modeling power of the C2f module is enhanced. For small targets, the P2 shallow multi-scale detection module SBA(Spatial Bi-directional attention) is proposed. The SBA module, a spatial bi-directional attention mechanism, is primarily applied to feature fusion in computer vision tasks. This module is designed to process multi-scale features, enabling effective integration between high-resolution and low-resolution features. Bidirectional pathways are established between high-and low-resolution features, while a P2 detection layer is incorporated to enhance shallow feature extraction capabilities. Through this approach, detection accuracy for objects at different scales is improved and spatial feature perception is strengthened. The RCS-YOLOv8 network was used to train and test the surface defect dataset of hot-rolled strip, and comparisons were made with the original model. Experimental results show that the average detection mean mAP and F1 score of the improved YOLOv8 algorithm on the NEU-DET dataset are 79.8% and 78%, respectively, which represent increases of 3% and 5% compared to the original model, and the floating-point operations of the improved algorithm are reduced by 2.4%. On the GC10-DET dataset, the mAP was increased by 4.2%. RCS-YOLOv8 not only achieves higher detection accuracy but also maintains a fast detection speed, which meets the requirements for both accuracy and real-time detection of industrial strip surface defects.更多还原