基于图像处理和深度迁移学习的芒果果实病状识别

目的 基于计算机层析成像(Computed tomography, CT)设备所得芒果CT序列图像，实现芒果内部品质的无损检测和病状识别分类。方法 利用分段函数法、中值滤波结合双边滤波，实现芒果图像增强；采用局部自适应阈值法，实现二值化处理；采用种子填充法进行区域填充；最后通过差影法准确提取芒果果实内部组织的坏损区域。基于深度迁移学习模型，对未处理和已处理的芒果图像数据开展训练和测试，通过AlexNet和GoogLeNet深度学习网络开展迁移学习，调整超参数完成训练过程的网络微调，在不同模型中对比未处理和已处理的芒果测试集在模型上的分类结果。结果 基于未处理数据集，GoogLeNet模型在学习率为0.0002下训练，Accuracy和Macro-average指标分别为98.79%和98.41%。基于已处理数据集，GoogLeNet模型在学习率为0.0002下训练，Accuracy和Macro-average指标分别为100%和100%。深度迁移学习模型在已处理数据集下的模型分类指标较未处理的数据集下有较大的提升。基于同一数据集且超参数一致时，GoogLeNet网络的分类效果明显优于AlexNet网络。结论 设定学习率为0.0002、迭代轮数为3、最小批值为64，基于GoogLeNet网络开展深度迁移学习训练，将所得模型作为最终的分类模型。

Recognition of mango fruit diseases based on image processing and deep transfer learning

Objective To realize non-destructive detection of the internal quality of mangoes and disease identification and classification based on the CT sequence images of mangoes obtained by computed tomography (CT) equipment.Method We used piecewise function method and median filter combined with bilateral filter to achieve image enhancement, used local adaptive threshold method to achieve binarization processing, used seed filling method to perform area filling, and used the image difference method to accurately extract the damaged area in inner tissue of mango fruit. Based on the deep transfer learning model, training and testing were carried out on unprocessed and processed mango image data. Transfer learning was carried out through the AlexNet and GoogLeNet deep learning networks, and hyperparameters were adjusted to complete the network fine-tuning of the training process. Under different models, the classification results of unprocessed and processed mango test sets on the model were compared.Result Based on the unprocessed data set, the GoogLeNet model was trained at a learning rate of 0.0002, and the Accuracy and Macro-average were 98.79% and 98.41% respectively. Based on the processed data set, the GoogLeNet model was trained at a learning rate of 0.0002, and the Accuracy and Macro-average were 100% and 100% respectively. The deep transfer learning model had a greater improvement in the model classification index of the processed data set than the unprocessed data set. Based on the same data set and consistent hyperparameters, the classification effect of the GoogLeNet network was significantly better than that of the AlexNet network.Conclusion While learning rate is set to 0.0002, the Epoch value is 3, and the Mini Batch value is 64, deep transfer learning training is carried out based on the GoogLeNet network, and the resulting model is used as the final classification model.