Apple Scab Detection Using Transfer Learning and Deep Convolutional Network

Every year, diseases and pests inflict enormous economic losses on the apple industry. One of the significant challenges faced by the farmers is the identification of the various diseases, as the signs and symptoms of several illnesses could be very similar and be present simultaneously. Through this proposed work, we make an effort to offer accurate and timely detection of one such apple disease, i.e., apple scab, a fungal disease. The main reason for choosing apple scab and apple leaves scab detection as our work is due to the lack of research on the topic, and we wanted to take this opportunity to do something meaningful and help the agriculture industry in India. The first part of the work was data preprocessing and labeling. The datasets containing photographs of patch-affected apples and leaves of an apple tree are collected; however, there are hardly any public datasets that contain enough images for us to use and train our models with because the acquisition of these images is extremely time-consuming and has a component of probability; therefore, we decided that transfer learning (TF) would be a suitable training approach. Training deep neural networks from scratch on a small dataset can take a long time and may not converge to a good solution due to overfitting. Transfer learning allowed us to start with a pretrained model and fine-tune it on your specific task. This significantly reduced the training time and resource requirements. The convolutional neural network (CNN) on the collected dataset is the model used to categorize apples. End-to-end learning algorithms known as CNN automatically extract characteristics from raw photographs and learn complex features from them. To avoid our model overfitting, we would use data augmentation techniques like rotation, translation, and scaling. An experimental result demonstrates that using the proposed structure of CNN and transfer learning, the results are comparatively better than the pretrained deep learning models.