INCREMENTAL LEARNING APPROACH FOR TOMATO LEAF DISEASE DETECTION WITHOUT CATASTROPHIC FORGETTING PROBLEM

Abstract

In the agricultural sector, the early detection on the crop disease is one of the major factor to prevent the diseases spread out and counteract the loss. However, in some cases the disease are still detected manually by the expert which is considered time-consuming, cost a lot of money, and somewhat inconsistencies occur. In the last decades, the utilization of machine learning has proven to allow the automation of identifying diseases on the leaf plant quickly and accurately. Nonetheless, the major problem has been faced as the lack of model to recognize the crop leaf diseases on the real condition. Therefore, huge number of various kind of leaf disease sample data is necessary to feed on the model. Incremental leaning is one of the best applicable approach to keep the model up to date by continually learning the new incoming plant leaf dataset. This study aims to classify the disease on the tomato leaves using CNN (Convolutional Neural Network) current state-of-the-art and implement the incremental learning as well as reducing the catastrophic problem by Freezing the last Layer and Rehearsal proposed method. The result shows that the best performance achieved when applying the Dense-Net with 95% accuracy and the proposed method succeed to outperform the highest previous performance on incremental learning which remain on the 94% of accuracy value after conducting incremental process on the base model