Desarrollo de una aplicación móvil para la detección de enfermedades en la mazorca de cacao utilizando deep learning.

Abstract

Cocoa is susceptible to many diseases that can cause devastating losses, this is one of the main problems for small and medium producers, and even for large ones. For this reason, it is key that cocoa crops carry out a correct phytosanitary control throughout their life cycle, mainly in the first months. Bad practices in crops greatly favor the appearance of diseases and, as a consequence, the loss of plantations. Environmental conditions can also influence their development when they are in a premature state. There are several disease prevention strategies, however, none of these measures ensures the complete elimination of the condition in cocoa. Therefore, it is necessary to adopt a control approach to manually identify the ears that have become affected. To solve this problem, the present work aims to develop a deep learning model capable of recognizing phytophthora and monilia diseases in cocoa pods. Due to the ability to identify several objects in an image, it was decided to choose an object detection algorithm such as YOLO in its most current version. The model will be incorporated into a mobile application as a final product, where images can be captured through the camera or gallery to execute the detection. The development process was based in the first instance on the construction of a dataset that contains images with sick and healthy ears. For this, the capture of photos and the addition of a dataset found in the Kaggle repository were considered. Subsequently, each image was labeled using LabelImg, selecting the ears with their respective classes in a bounding box. Then, the Roboflow platform was used, which allowed us to automate the tasks so that our dataset has the desired format. The training of the model with YOLOv5 was carried out in the cloud, using the Google Collaboratory service to make use of a GPU. The training was based on the transfer learning technique to improve our final model and obtain better results given the limitation of our dataset. The final model obtained values of 0.80 for mAP, 0.88 for precision and 0.73 for recall. To use it in a mobile application, the model was converted to an ONNX format and then with the Tencent NCNN framework to its respective format. The implementation of the mobile application was carried out according to the procedures established in the Mobile-D methodology, so the phases were worked with iterations. The application was developed using Android Studio and was made up of two modules: detection and gallery. Thus allowing the execution of the model within the detection module and the display of the resulting images in the gallery module. To assess the performance of our model, the tests were performed using the test set in the mobile application. From the tests, a confusion matrix was obtained for each class and the following metrics: precision of 0.85, recall of 0.84 and f-score of 0.84. With these results, it is validated that our model, together with the application, is capable of detecting cobs affected by phytophthora and monilia, in addition to those that are not affected.