基于改进YOLOv5s的铝型材表面弱缺陷识别方法

Weak Defect Identification Method on Profile SurfaceBased on Image Enhancement and YOLOv5s

针对采用传统的机器视觉方法识别铝型材表面弱缺陷存在的效率低和精度差的问题,提出一种将数据增强与YOLOv5s相结合的铝型材表面弱缺陷识别方法。采用anchor-free方法简化人工设计YOLOv5参数的步骤,降低检测复杂度;利用解耦检测器解决YOLOv5s检测中分类与回归任务冲突的问题,加快损失函数的收敛速度。通过优化算法的边界框回归损失函数,提高算法模型的定位精度;同时引入γ参数解决弱缺陷样本不平衡的问题。通过图像马赛克与像素混合方法提升模型对弱缺陷图像的识别能力。试验结果表明,改进算法的检测平均精度均值为93.3%,检测速度为41帧/秒,能提高船舶类铝型材弱缺陷检测的效率和自动化程度。

Aiming at the problems of low efficiency and poor accuracy in identifying weak defects,a method for combining data enhancement and YOLOv5s was proposed.The anchor-free method is adopted to simplify the steps of manual design of YOLOv5s parameters and reduce the detection complexity.The decoupling detector is used to solve the problem of conflict between the classification and regression tasks in the YOLOv5s detection and accelerate the convergence of the loss function.By optimizing the bounding box loss function of the algorithm,the positioning accuracy of the algorithm model is improved;at the same timeγParameter is introduced to solve the problem of unbalanced weak defect samples.The image Mosaic and pixel hybrid method is used to improve the recognition ability of the model for weak defect images.The result shows that the average detection accuracy of the improved algorithm is 93.3%,and the detection speed is 41 FPS,which improves the efficiency and automation of weak defect detection of ship aluminum profiles.