基于改进目标检测算法的道路场景图像实例分割方法  

An Instance Segmentation Method for Road Scene Images Based on Improved YOLOv5

作　　者：周洲 Zhou Zhou(College of Optoelectronic Information and Computer Engineering,University of Shanghai for Science and Technology,Shanghai)

机构地区：[1]上海理工大学光电信息与计算机工程学院,上海

出　　处：《建模与仿真》2025年第3期281-293,共13页Modeling and Simulation

摘　　要：实例分割是图像分割的重要组成部分,同时也是计算机视觉中的一个关键研究课题,广泛应用于自动驾驶和安全监控等领域。然而,由于道路场景通常具有复杂性、多样性和杂乱的特点,处理这些场景变得尤为挑战。针对道路场景图像实例分割难度大、精度低、定位不精确等问题,本文提出一种基于改进YOLOv5(You Only Look Once version 5)的道路场景实例分割算法。以YOLOv5为基础模型,在Head模块中采用RFAConv(Receptive-Field Attention Convolution)卷积代替部分传统卷积,它全面解决了卷积核的参数共享问题,考虑到接受域中每个特征的重要性,提供了几乎可以忽略不计的计算成本和参数增量,能够更好地捕捉和融合图像特征,提升分割的精度和鲁棒性。采用ShapeIOU代替YOLOv5中原损失函数CIOU(Complete-IoU),通过聚焦边框自身形状与自身尺度计算损失,使得边框回归更为精确,能够有效提升检测效果且优于现有方法。实验结果表明:与原模型相比,改进后的模型的分割精度mAP50(mean Average Precision)达到了33.8%,相较于YOLOv5s,优化后的模型在分割精度上提高了1.2%,能够更加高效地完成道路场景的图像分割任务。Instance segmentation is a crucial component of image segmentation and serves as a significant re-search area within computer vision.It finds extensive applications in various domains,including autonomous driving and security surveillance.However,due to the complexity,diversity,and clut-tered nature of road scenes,handling these scenarios becomes particularly challenging.In response to the challenges of high complexity,low accuracy,and imprecise positioning in road scene image instance segmentation,this paper introduces an enhanced YOLOv5-based(You Only Look Once ver-sion 5)algorithm specifically designed for road scene instance segmentation.Taking YOLOv5 as the base model,RFAConv(Receptive-Field Attention Convolution)convolution is used in the Head mod-ule instead of part of the traditional convolution,which comprehensively solves the parameter shar-ing problem of convolution kernel,considers the importance of each feature in the receptive do-main,provides almost negligible computational cost and parameter increment,and is able to better capture and fuse the image features to improve the segmentation accuracy and robustness.ShapeIOU is used instead of the original loss function CIOU(Complete-IoU)in YOLOv5,and the loss is calcu-lated by focusing on the shape of the frame itself and the scale of the frame itself,which makes the frame regression more accurate,and it can effectively improve the detection effect and outperform the existing methods.The experimental results show that compared with the original model,the segmentation accuracy mAP50 of the improved model reaches 33.8%,and compared with YOLOv5s,the optimised model improves the segmentation accuracy by 1.2%,which is able to complete the image segmentation task of the road scene more efficiently.

关 键 词：实例分割 YOLOv5 接受场注意力卷积 注意力机制 损失函数 特征融合 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]