基于卷积神经网络和迁移学习的番茄成熟度检测

Tomato maturity detection based on convolutional neural network and transfer learning

为解决现有番茄成熟度检测模型存在的计算复杂且精度低的问题,以普罗旺斯番茄为样品,利用HSV通道数据进行分析,将样本图片划分为绿熟期、变色期和红熟期,建立不同成熟度番茄的数据集。基于迁移学习,对MobilenetV2、InceptionV3和VGG19三种卷积神经网络模型的超参数进行了优化,建立了基于番茄成熟度的检测模型。进一步分析了SGD、Adam、Adagrad优化算法和不同训练轮数的检测效果,发现MobilenetV2的表现最好,在优化器为Adam、训练100轮时准确率可达99.05%,损失值为0.038 4,满足高精度检测。检测方法能在低损失值下获得更高的检测准确率,可为番茄成熟度检测研究提供参考。

In order to solve the problems of complicated calculation and low accuracy of existing tomato ripener detection models,this paper takes Provence tomatoes as samples and conducts analysis using HSV channel data.The sample images are divided into green ripening,discoloration,and red ripening stages,and a dataset of tomatoes with different maturity levels is established.Based on transfer learning,the hyper parameters of three convolutional neural network models,MobilenetV2,InceptionV3,and VGG19,were optimized,and a detection model based on tomato maturity was established.The detection effects of SGD,Adam and Adagrad optimization algorithms and different training rounds were further analyzed,and it was found that MobilenetV2 had the best performance,with an accuracy of 99.05%and a loss value of 0.0384 when the optimizer was Adam and 100 training rounds,meeting the high efficiency maturity detection.The detection method can obtain higher detection accuracy under low loss value,and provide reference for the detection of tomato maturity.