A Super-Efficient and Compact CNN for Tomato Leaf Disease Detection on Resource-Constrained Devices

Abstract

The diseases that affect tomato leaves have a major impact on the agricultural industry. Thus, it is crucial to develop an automated detection system that is both accurate and lightweight. In this context, the implementation of lightweight CNN architecture using the concept of depth-wise separable convolution and squeeze-excitation has been used for the purpose of accurate classification of diseases that affect the tomato leaves. The results that are obtained using the proposed CNN architecture indicate that the classification of diseases that affect the tomato leaves is done in a more accurate manner. In this context, it is observed that the accuracy for the training set is 100%, the loss is 0.0217, the accuracy for the validation set is 99.83%, the loss is 0.0263, and the accuracy for the test set is 99.75%. Thus, it is observed that a high AUC score is obtained, which is 0.999996. The proposed CNN architecture has a total of 219,651 trainable parameters. In this context, it is observed that the total parameters are of a size of 858.01 KB.