摘要
在遮挡目标识别中,目标可能会被其他物体遮挡,导致目标的部分有效特征丢失或变形。目标有效特征的减少,使得单一YOLOv4(You Only Look Once version 4)无法准确识别锚框的初始值,使得模型目标识别困难。为此,引入K-means++算法改进单一YOLOv4算法,提出基于改进YOLOv4的遮挡目标识别算法。通过非下采样Contourlet变换划分图像为低频部分和高频部分,分别利用线性增强函数和改进的自适应阈值增强图像,并经由非下采样Contourlet逆变换生成重建图像,对其执行模糊对比度增强。选取YOLOv4作为目标识别基础模型,采用深度可分离卷积替代模型中部分卷积,并替换特征金字塔为递归特征金字塔,提升小目标和遮挡目标的特征学习能力。引入K-means++算法自适应获取锚框,优化锚框初始值,并利用完全交并比和交叉熵构建损失函数,训练改进的YOLOv4并将增强后图像输入其中,实现遮挡目标识别。实验结果表明,所提方法能够有效识别图像目标,且识别结果P-R曲线更理想。
In occlusion target recognition,the target may be occluded by other objects,resulting in the loss or deformation of some effective features of the target.The reduction of effective features of the target makes the single YOLOv4(You Only Look Once version 4)unable to accurately recognize the initial value of the anchor box,making the model target difficult to recognize.For this reason,K-means++algorithm was introduced to improve the single YOLOv4 algorithm,and a occlusion target recognition algorithm based on the improved YOLOv4 was proposed.The image was divided into low frequency and high frequency parts by nonsubsampled Contourlet transform.The linear enhancement function and improved adaptive threshold were used to enhance the image,and the reconstructed image was generated by non-subsampled Contourlet inverse transformation,which performed the fuzzy contrast enhancement.YOLOv4 was selected as the basic model of target recognition,and the depth-separable convolution was used to replace part of the convolution in the model,and the feature pyramid was replaced with a recursive feature pyramid to improve the feature learning ability of small targets and occluded targets.The K-Means++algorithm was introduced to adaptively obtain the anchor frame,optimize the initial value of the anchor frame,and construct the loss function by using the complete intersection ratio and cross entropy,and train the improved YOLOv4 and input the enhanced image into it to realize the occlusion target recognition.The experimental results showed that the proposed method can effectively recognize image targets and the P-R curve of recognition results is more ideal.
作者
裴云霞
黄忠
PEI Yunxia;HUANG Zhong(School of Information and Finance,Xuancheng Vocational&Technical College,Xuancheng Anhui 242000,China;School of Electronic Engineering and Intelligent Manufacturing,Anqing Normal University,Anqing Anhui 246133,China)
出处
《海南热带海洋学院学报》
2025年第2期106-113,共8页
Journal of Hainan Tropical Ocean University
基金
安徽省高等学校科学研究项目(2022AH052783)。
作者简介
裴云霞,女,安徽宣城人,副教授,硕士,研究方向为计算机图形图像;黄忠,男,安徽岳西人,教授,博士,研究方向为机器视觉、人机交互。
