CountNet:一种用于堆叠胶合板计数的自监督学习框架  

CountNet:A Self-Supervised Learning Framework for Automated Stacked Plywood Counting

作　　者：苏凡 王若琪 王海涛[1] SU Fan;WANG Ruoqi;WANG Haitao(Department of Computer Science and Engineering,Sun Yat-sen University,Guangzhou 510006,China;Department of Computer Science and Engineering,Hong Kong University of Science and Technology,Guangzhou 511453,China)

机构地区：[1]中山大学计算机学院,广东广州510006 [2]香港科技大学计算机学院,广东广州511453

出　　处：《软件导刊》2025年第2期19-25,共7页Software Guide

基　　金：广东省自然科学基金项目(2021A1515011319)。

摘　　要：自动化胶合板计数是工业生产中的一大难题,传统基于人工计数和物理计数的方法耗时且低效。然而,堆叠胶合板图像又存在着边缘不均匀、厚度不规律等干扰因素,现有的深度学习算法提取到的特征代表性不强,导致计数结果不准确。针对上述问题,提出一种用于堆叠胶合板计数的自监督学习框架——CountNet。CountNet针对计数问题优化了损失函数的使用方式,该损失函数利用对比学习的方法,可以进一步放大正负样本间的差异,使网络能提取到更具代表性的视觉特征。最后将该特征投入下游任务中,完成计数。实验结果表明,该方法在准确率、损失下降等方面均优于其他常见的计数模型,证明了其在计数能力上的优越性。Automated counting of stacked plywood materials is a major challenge in industrial production.Traditional methods based on manual counting and physical counting are time-consuming and inefficient.However,stacked plywood images are often affected by factors such as uneven edges and irregular thickness,leading to inaccurate counting results with existing deep learning algorithms due to the lack of strong representational features extracted.To address these issues,we propose a self-supervised learning framework,CountNet,for counting stacked plywood materials.CountNet introduces a novel loss function that leverages the advantages of contrastive learning to further amplify the differences between positive and negative samples,enabling the network to extract more representative visual features.These features are then utilized in downstream tasks to achieve accurate counting.Experimental results demonstrate that the proposed method outperforms other common counting models in terms of accuracy,loss reduction,and various other metrics,showcasing its superiority in counting capability.

关 键 词：自监督对比学习 计算机视觉 堆叠胶合板计数 数据增强技术 损失函数优化 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]