融合自适应下采样的带钢表面缺陷检测算法

Fusion of adaptive down-ssampling for strip steel surface defect

针对当前带钢表面缺陷检测领域,因终端检测设备计算资源受限,导致基于深度学习的网络模型在终端检测速度降低的问题,提出一种融合自适应下采样的带钢表面缺陷检测算法。首先构建自适应下采样模块(ADWO),通过引入权重系数实现网络对不同特征的关注,且分组卷积的结构设计降低模块的参数,从而实现提高网络对特征感知能力的同时降低网络复杂度;其次,结合context guided block构建CG-C3模块,通过联合局部和周围上下文信息来提升网络对缺陷特征的提取与整合能力;最后,为解决带钢表面微小缺陷难以检测的问题,在网络中引入RFB(Receptive Field Block)注意力机制,通过整合多尺度缺陷特征,提升网络的检测精度。实验结果表明,在NEU-DET数据集上,改进后的算法平均检测精度达到79.5%,较于原YOLOv5s网络提升2%且FPS提升了6帧/s,参数量降低39.9%,浮点运算量降低39.8%。通过消融实验和对比实验验证了各个改进模块的有效性和所提算法的优越性。

To tackle the issue of reduced speed in deep learning-based network models due to limited computational resources in terminal detection equipment.An algorithm for detecting surface defects on strip steel was proposed,incorporating adaptive down-sampling.They first constructed the adaptive down-sampling module(ADWO)to enhance the network's attention to different features by introducing weight coefficients and reducing module parameters through group convolution.Then,the CG-C3 module was designed,combined with the context guided block,to improve the network's ability to detect defective features by integrating local and surrounding context information.Additionally,the RFB attention mechanism was introduced to enhance the detection accuracy of small defects on the strip surface by integrating multiscale defect features.Experimental results showed that the improved algorithm achieved an average detection accuracy of 79.5%on the NEU-DET dataset,surpassing the original YOLOv5s network by 2%.Moreover,it increased FPS by 6 frames/s,reduced the number of parameters by 39.9%,and decreased floating-point operations by 39.8%.Ablation and comparison experiments confirmed the effectiveness of each improvement module and the superiority of the proposed algorithm.