基于GAF与改进CNN的苹果多品质指标无损检测方法研究

Research on Non-destructive Detection Method for Apple Multiple Quality Indicators Based on GAF and Improved CNN

摘要

本研究以4个品种共600个无损伤健康苹果为试验材料，以近红外光谱数据为基础，构建了一种基于格拉姆角场（GAF）与SE通道注意力机制的双分支卷积神经网络融合模型（GAF-SE-1D-2D-CNN），实现了苹果的硬度、糖度、可滴定酸度三项核心品质指标的同步无损检测。共采集了600组原始近红外光谱数据，经正态分布3σ准则等异常值检验后，获得有效样本591个，经过对比SG平滑+SNV、SG平滑+MSC、SG平滑+一阶导数三种光谱预处理方案的优化效果，采用KS算法按照7∶3比例将有效样本划分为训练集与测试集，再通过GAF结合SE通道注意力机制进而强化品质相关特征通道，构建出了双分支卷积神经网络完成多任务的回归预测。模型训练采用Adam优化器，配套L2正则化、Dropout、学习率衰减与早停策略抑制过拟合，以决定系数和均方根误差为评价指标。最终结果表明，该模型在独立测试集上表现优异，硬度、糖度、可滴定酸度的 R²分别达到0.875 5、0.880 2、0.914 1，RMSE分别为0.16 10、0.508 1、0.020 5，三项指标预测精度均满足果品无损检测需求。

Abstract

This study utilizes 600 intact and healthy apples from four varieties as experimental materials. Based on near- infrared spectroscopy data, a dual-branch convolutional neural network fusion model (GAF-SE-1D-2D-CNN) is constructed, which incorporates Gramm's angular field (GAF) and SE channel attention mechanism. This model achieves synchronous non-destructive detection of three core quality indicators of apples: firmness, sugar content, and titratable acidity. A total of 600 sets of original NIR spectral data are collected, and after outlier detection using the 3σ criterion of normal distribution, 591 valid samples are obtained. After comparing the optimization effects of SG smoothing + SNV, SG smoothing + MSC, SG smoothing + first-order derivative, the valid samples are divided into a training set and a test set in a 7∶3 ratio using the KS algorithm. The GAF combined with the SE-channel attention mechanism is then used to enhance the quality-related feature channels, and the dual-branch convolutional neural network is constructed to complete multi-task regression prediction. The model training employs Adam optimizer, along with L2 regularization, dropout, learning rate attenuation and early stop strategy to suppress overfitting, with the decision coefficient and root mean square error as evaluation indicators. The final results show that the model performs well on the independent test set, with the R² of firmness, sugar content and titratable acidity reaching 0.875 5, 0.880 2 and 0.914 1, respectively, and the RMSE of 0.161 0, 0.508 1 and 0.020 5, respectively. The prediction accuracy of all three indicators meet the requirements for non-destructive fruit quality detection.

关键词

近红外光谱 / 苹果品质 / GAF / CNN / 无损检测

Key words

NIR Spectroscopy / apple quality / GAF / CNN / non-destructive detection

参考文献

[1] 王嘉琪,张 洁, 任玉锋, 等. 宁夏不同地域产区嘎啦系苹果果实品质比较分析[J]. 经济林研究, 2025, 43(03): 242-251.

[2] 曾小松, 宦克为, 刘小溪, 等. 基于竞争性自适应重加权采样和格拉姆角差场的二维卷积神经网络近红外光谱分析模型研究[J]. 分析化学, 2025, 53(06): 955-966.

[3] 毕钲发,刘 杨, 伍新明, 等. 基于深度学习的海底线缆大规模快速普查[J]. 地球物理学报, 2025, 68(12): 4881-4894.

[4] Liu J, Wu J, Xie H, et al. Semantic segmentation of urban remote sensing images based on deep learning[J]. Applied Sciences, 2024, 14(17): 7562.

[5] Brahim B, Ginés M G, Sajad S, et al. Convolutional neural networks for estimating the ripening state of Fuji apples using visible and near-infrared spectroscopy[J]. Food and Bioprocess Technology, 2022, 15(10): 2226-2236.

[6] Oliveira-Folador G, Bicudo O D M, Andrade D F E, et al. Quality traits prediction of the passion fruit pulp using NIR and MIR spectroscopy[J]. LWT, 2018, 95: 172-178.

[7] Alessio S, Glenn M O. Relationships between soluble solids and dry matter in the flesh of stone fruit at harvest[J]. Analytica, 2021, 2(1): 14-24.

[8] Rong D, Wang H, Ying Y, et al. Peach variety detection using VIS-NIR spectroscopy and deep learning[J]. Computers and Electronics in Agriculture, 2020, 175: 105553.

[9] 刘燕德, 王军政, 姜小刚, 等. 多模式可调节光学机构的苹果可溶性固形物近红外光谱检测[J]. 光谱学与光谱分析, 2021, 41(07): 2064-2070.

[10] 雷 鹰, 刘翠玲, 周子彦. 应用便携式近红外光谱仪研究苹果糖度的快速分析模型[J]. 食品科学技术学报, 2018, 36(06): 95-100.

[11] 徐淑彦, 王 伟, 李世雄. 基于近红外光谱的苹果内部糖度快速检测方法[J]. 中国农机装备, 2025(11): 22-25.

[12] 蔡健荣, 黄楚钧, 马立鑫, 等. 一维卷积神经网络的手持式可见/近红外柑橘可溶性固形物含量无损检测系统[J]. 光谱学与光谱分析, 2023, 43(09): 2792-2798.

[13] 张田煜, 张晓蕾, 刘少华, 等. 基于1D-CNN-RF 的梨果可见/近红外糖度预测模型[J]. 南京农业大学学报, 2026, 49(02): 433-441.

[14] 邓 娅, 王新越,黄 洋, 等. 基于近红外光谱与格拉姆角场多模态融合的水果品质分类模型[J]. 物联网技术, 2026, 16(03): 116-122.

[15] 樊书祥, 王庆艳, 杨雨森, 等. 水果糖度可见-近红外光谱手持式检测装置开发与试验[J]. 光谱学与光谱分析, 2021, 41(10): 3058-3063.

[16] 徐惠荣, 李青青. 皇冠梨糖度可见/近红外光谱在线检测模型传递研究[J]. 农业机械学报, 2017, 48(09): 312-317.

[17] 刘鸿阳, 孔德国, 罗华平, 等. 基于多光谱图像角度融合测定库尔勒香梨理化指标[J]. 光谱学与光谱分析, 2024, 44(03): 649-655.

[18] 陈建新. 基于近红外光谱的苹果硬度便携式检测设备研究[D]. 杨凌: 西北农林科技大学, 2018.

[19] 杨光辉, 张永立, 王美蟠, 等. 近红外光谱检测鲜食玉米可溶性固形物含量[J]. 光谱学与光谱分析, 2025, 45(07): 1932-1939.

[20] 孙伟哲,李 含, 陈希源, 等. 电力电缆绝缘材料型号的近红外光谱快速识别方法研究[J]. 光谱学与光谱分析, 2025, 45(11): 3153-3159.

[21] Meng L, Chen G, Liu D, et al. Universal modeling for non-destructive testing of soluble solids content in multi-variety blueberries based on hyperspectral imaging technology[J]. Applied Sciences, 2025, 15(7): 3128.

[22] 董诗佳. 基于近红外光谱技术的梨成熟度判别模型构建[D]. 杨凌: 西北农林科技大学, 2025.

基金资助

山东省属普通本科高校教师访学研修经费资助(鲁教师函（2025〕33 号)