融合少样本学习与注意力端到端网络的小目标在线检测研究  

Research on Online Detection of Small Foreign Objects Using Few-Shot Learning and Attention-Based End-to-End Networks

作　　者：周庆辉 葛馨远 孙峥 陈盛开 周煜 ZHOU Qinghui;GE Xinyuan;SUN Zheng;CHEN Shengkai;ZHOU Yu(Guangzhou Baiyun Power Supply Bureau of Guangdong Power Grid Co.,Ltd.,Guangzhou Guangdong 550014,China)

机构地区：[1]广东电网有限责任公司广州白云供电局,广东广州550014

出　　处：《机床与液压》2024年第17期130-135,共6页Machine Tool & Hydraulics

基　　金：南方电网公司科技项目(082300KK52210003)。

摘　　要：小目标检测是计算机视觉领域的研究方向之一,旨在解决在图像或视频中检测和定位尺寸较小的目标的问题。由于小目标往往具有低分辨率、模糊、被遮挡等特点,传统的目标检测算法在处理小目标时存在挑战。对此,提出一种融合少样本学习与注意力端到端网络的小目标检测方法。该方法通过引入图像增强技术和注意力机制,对传统的端到端检测网络进行优化,以提高检测性能。通过数据增强方式,对原始数据进行扩充,增加数据的多样性和数量;引入注意力机制,提取图像中的关键信息,以提升检测结果的准确性;最后,在网络结构方面,将原有的特征金字塔网络(FPN)替换为加权双向特征金字塔网络(BiFPN),以获取更丰富的图像特征。实验结果表明:通过图像增强和注意力机制,所提方法的精准率、召回率、平均精度均值和检测速度在训练尺度为640像素×640像素时分别为98.41%、99.54%、99.50%和28帧/s,相较于YOLOv5算法分别提升了2.91%、5.9%、1.93%和2帧/s,验证了该方法的有效性和可行性。Small object detection is one of the research directions in the field of computer vision,aiming to address the problem of detecting and locating small objects in images or videos.Traditional object detection algorithms face challenges in handling small objects due to their low resolution,blurriness,and occlusion.To address this issue,a novel approach was proposed for small foreign object detection that combined few-shot learning with attention-based end-to-end networks.In this method,the traditional end-to-end detection network was optimized by introducing image enhancement and attention mechanisms to improve the detection performance.The original data were augmented using data augmentation techniques to increase the diversity and quantity of the data.An attention mechanism was incorporated to extract crucial information from the images and improve the accuracy of the detection results.In terms of network structure,the original feature pyramid network(FPN)was replaced with bi-directional feature pyramid network(BiFPN)to obtain richer image features.Experimental results demonstrate that when the training scale is 640 pixels×640 pixels,the proposed method achieves precision,recall,mean average precision(mAP),and detection speed of 98.41%,99.54%,99.50%,and 28FPS,respectively,surpassing mainstream algorithms by improvements of 2.91%,5.9%,1.93%and 2FPS through image augmentation and attention mechanisms.The effectiveness and feasibility of the proposed method are validated.

关 键 词：小目标在线检测 深度学习 图像增强技术 注意力机制 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]