An empirical performance comparison of deep learning models for apple leaf disease recognition

Currently, Biotic and abiotic factors, in agriculture, are estimated to be responsible for around 22% of the total losses in crop production. It is therefore very important, but at the same time very difficult for farmers to identify these constraints at the earliest possible stage. The most recent developments in computer vision have made it possible to identify as well as classify plant lesions at a primary level. Various deep learning techniques have made it possible to detect biotic stress in apple plant leaves with a very high degree of accuracy. Hence, automated segmentation of images of diseased leaves has become a leading research theme all over the world. In this paper, a dataset from FGVC7, which is an open, source dataset and contains four different classes of leaf diseases of the apple crop is used to measure the performance of the proposed model. A three, step image preprocessing procedure is conducted, which involves image enhancement, removal of background noise, and object identification through a variety of approaches. Using UNet, PSPNet, and UNet++, disease areas are identified. UNet + + delivering the most precise object segmentation. The segmented patch data are then fed into a fusion, based classification model as input. Patch segmentation makes it possible to separate the various disease areas, thus increasing the classification accuracy. The proposed framework combines several CNN architectures, namely, VGG16, InceptionV3, AlexNet, and GoogleNet, in order to take advantage of their complementary strengths, then a Vision Transformer (ViT) is used to identify distinct local features via early feature fusion. After this, the efficiency of the proposed model is check in comparison with the pre-trained CNN models individually. Comparing with all other models the proposed model has shown the accuracy of highest order i.e. 98.85%, clearly out performing individual CNN models.