基于注意力机制的水果新鲜度检测可解释模型  

Fruit Freshness Detection Explainable Model Based on Attention Mechanism

作　　者：张寅升[1] 宋曾林 王海燕[1] Zhang Yinsheng;Song Zenglin;Wang Haiyan(Zhejiang Food and Drug Quality&Safety Engineering Research Institute,Zhejiang Gongshang University,Hangzhou 3100182School of Management and E-Business,Zhejiang Gongshang University,Hangzhou 310018)

机构地区：[1]浙江工商大学食药质量安全工程研究院,杭州310018 [2]浙江工商大学管理工程与电子商务学院,杭州310018

出　　处：《中国食品学报》2024年第10期28-36,共9页Journal of Chinese Institute Of Food Science and Technology

基　　金：国家重点研发计划项目(2023YFD1000400);国家自然科学基金项目(62376249,91746202)。

摘　　要：近年来,基于机器视觉和深度学习的水果新鲜度检测成为主流方法之一。针对现有的深度学习技术,对卷积神经网络在水果的特征提取方面的应用进行探讨,在ResNet34主干网络中引入卷积注意力模块(CBAM),以实现水果新鲜度的检测,并采用类激活图(CAM)系列技术对于能够反映图片关键特征的像素进行热力图可视化。在水果公开数据集上,引入注意力机制前、后的ResNet34分类准确率分别为96.80%和99.71%。同时,CAM热力图反映注意力模型能够更加准确地捕获水果图像中变质腐烂的区域,表明提出的模型改善了深度学习特征提取的能力,不仅提高了模型的泛化能力,而且增强了模型的可解释性。In recent years,fruit freshness detection based on machine vision and deep learning has become one of the mainstream methods.This study explores the application of deep learning technologies,particularly convolutional neural networks(CNNs),in feature extraction for fruit freshness identification.This paper introduces the CBAM(Convolutional block attention module)attention mechanism module into the ResNet34(34-layer residual network)backbone network to achieve fruit freshness detection.Class activation mapping(CAM)techniques are employed to visualize the heatmaps of pixels that reflect the critical features in the images.On a public fruit dataset,the classification accuracy of the ResNet34 network before and after introducing the attention mechanism is 96.80%and 99.71%,respectively.The CAM heatmaps show that the attention model can more accurately capture the regions of interest in the fruit images,indicating that the proposed model improves the feature extraction capability of deep learning,not only enhancing the model's generalization ability but also increasing its interpretability.

关 键 词：水果新鲜度检测 深度学习 注意力机制 残差网络 

分 类 号：TS255.7[轻工技术与工程—农产品加工及贮藏工程] TP18[轻工技术与工程—食品科学与工程] TP391.41[自动化与计算机技术—控制理论与控制工程]