摘要
在锂离子电池顶盖表面缺陷检测时,复杂光照条件下,对不规则小目标检测的准确率低,且面临轻量化与实时性挑战。提出一种电池顶盖缺陷检测(BCDD)-YOLO算法。首先,引入部分卷积代替常规卷积,构造轻量化的逐点卷积和部分卷积相互合作混洗的混合卷积PPW-Conv模块,减少计算冗余和内存访问;其次,在特征融合部分,提出线性可变性的大核注意力LD-LKA模块,结合大卷积核和线性可变形卷积的注意力机制,广泛灵活地捕捉图像的上下文信息;最后,使用更有效聚焦于样本的损失函数InnerFocaler-CIoU代替完全交并比CIoU损失函数,使用辅助边界框加速边界框的回归,提高小目标的检测精度和鲁棒性。实验表明,该方法与主流算法相比检测精度明显提升,在锂离子电池顶盖缺陷数据集上的平均精度均值mAP达到了75.5%,充分验证了该改进算法的有效性。
When detecting the surface defects of Li-ion battery top cover,the detection accuracy for irregular small targets is low under complex lighting conditions,and there are challenges in terms of lightweighting and real-time performance.A battery cover defect detection(BCDD)-YOLO algorithm is proposed.Firstly,partial convolution is introduced instead of the conventional convolution to construct a lightweight point-wise convolution and partial convolution cooperative shuffling mixed convolution module PPW-Conv,which reduces computational redundancy and memory access;Secondly,in the feature fusion part,a linearly deformable large kernel attention module LD-LKA is proposed,combining the attention mechanism of large convolution kernels and linear deformable convolutions,to widely and flexibly capture the context information of the image;Finally,a more effective loss function InnerFocaler-CIoU that focuses more on the samples is used instead of the complete intersection-over-union CIoU loss function,an auxiliary bounding box is used to accelerate the regression of bounding boxes,improving the detection accuracy and robustness of small targets.Experiments show that this method significantly improves the detection accuracy compared to mainstream algorithms,the average precision mean mAP on the Li-ion battery top cover defect dataset reaches 75.5%,fully verifying the effectiveness of the improved algorithm.
作者
刘志辉
曹丽丽
朱勇建
LIU Zhihui;CAO Lili;ZHU Yongjian(School of Mechanical and Energy Engineering,Zhejiang University of Science and Technology,Hangzhou 310023,Zhejiang,China;Ningbo Minjie Information Technology Co.,Ltd.,Ningbo 315300,Zhejiang,China)
出处
《电池》
北大核心
2025年第6期1248-1256,共9页
Battery Bimonthly
基金
教育部中国高校产学研创新基金(2024HY010)
浙江省“十四五”教学改革项目(jg20220405)。
