Improved detection and identification approach in tomato leaf disease using transformation and combination of transfer learning features

Research in Artificial Intelligence has helped to achieve significant progress during the last decade, in various sectors. As concern agriculture, tomato is one of the most cultivated and consumed vegetables in the world because of its nutritional and therapeutic properties. However, this crop faces phytosanitary problems. This study contributes to the automatic Disease Detection and Identification (DDI) in tomato plants using neural networks. The proposed model performs the DDI of tomato from the images of tomato leaves or the videos from cameras installed in the plantation. The extraction of the tomato leaves images in the video was done using object recognition by RGB color. We extract the features from the images using transfer learning based on the pre-trained ResNet101 and ResNet152 models, and their classification is done by the multi-layer perceptron. To obtain better accuracy, the characteristics of our images were extracted at several levels from the pre-trained model. We observed that the recognition rate increased with the number of layers except for the last layer for which the accuracy was low. It was also observed that the replacement of the input vectors by the distance from the mean improved the results obtained in the literature. As a result of this work, we obtained an accuracy of 97.26% and 97.5% from transfer learning of the two pre-trained architectures and we improve these state-of-art results to 98.3% using fivefold with the concatenation of two mean deviation features extracted from ResNet101 and ResNet152. In terms of maximum accuracy, we obtained 98.9%. These results will be used to calibrate the automatic mobile camera system moving on rails.