Flooding-based MobileNet to identify cucumber diseases from leaf images in natural scenes

Cucumber production in China is declining due to various pathologic diseases, but the technology for plant disease detection is not mature and requires high labor costs. Moreover, since planting sites are typically high-density scenes, most photos are shot from various angles with messy backgrounds, resulting in poor classification reliability. In this paper, batches of cucumber leaf image data are collected from agricultural websites and then preprocessed through the image size normalization. A mobile-based recognition algorithm is proposed to identify cucumber diseases from leaf images in natural scenes, enabling farmers to detect diseases more quickly. The proposed algorithm allows farmers to upload cucumber pictures, and rapidly and accurately classify them with high accuracy. With a improved network based on MobileNet V3, the classification of seven kinds of cucumber leaf diseases can be quickly and accurately completed. The network model is achieved by selecting appropriate parameters, optimizers, and batch capacity using the single-variable method. Additionally, a new training strategy called the flooding method is applied in the model, replacing the traditional strategy that relies solely on loss decline. An accuracy of 83.3% is achieved on our custom dataset. Finally, two public datasets, namely PlantVillage and Apple Disease, are selected for migration experiments. The achieved accuracy rates for these datasets are 99% and 98.1% respectively, demonstrating the universality of the proposed strategy. The code for all the experiments will be made available for reference on the GitHub repository at https://github.com/YiQuanMarx/Agricultural_Diseases_Dentification.