ResdenseNet: a lightweight dense ResNet enhanced with depthwise separable convolutions and its applications for early plant disease classification

作     者：Nagpal, Jyoti Goel, Lavika 

作者机构：Department of Computer Science & Engineering The NorthCap University Gurugram122017 India Department of Computer Science and Engineering Malaviya National Institute of Technology Rajasthan Jaipur302017 India 

出 版 物：《Neural Computing and Applications》 (Neural Comput. Appl.)

年 卷 期：2025年第37卷第8期

页      面：6305-6326页

核心收录：

学科分类：0710[理学-生物学] 0810[工学-信息与通信工程] 08[工学] 0812[工学-计算机科学与技术（可授工学、理学学位）] 

基　　金：The authors declare that no funds  grants  or other support were received during the preparation of this manuscript 

主　　题：Plant diseases 

摘      要：In recent years, artificial intelligence has undergone robust development, leading to the emergence of numerous autonomous AI applications. However, a crucial challenge lies in optimizing computational efficiency and reducing training time while maintaining high accuracy with limited hardware resources. This paper introduces ResdenseNet, a model built upon the MobileNet, DenseNet, and ResNet architectures. ResdenseNet combines dense blocks and residual blocks from the DenseNet and ResNet architectures. In these dense blocks, the standard convolutional units are replaced by depthwise separable convolutional units, a significant part of the MobileNet architecture. The experimental outcomes are contrasted with established models and their iterations, including ResNet-50, ResNet-101, MobileNet-V1, MobileNet-V2, DenseNet-121, and DenseNet-169. The proposed model is tested on benchmark and proposed datasets, showcasing its efficiency in reducing computations and accelerating the training process. Emphasizing hyperparameter importance, ResdenseNet, optimized with a growth rate of 64, 6 layers, and ReLU activation, achieves an accuracy of (98.73%) and a F1-score of (98.20%) on the wheat and barley dataset. The results indicate that ResdenseNet significantly decreases the number of parameters to 0.72M and efficiently shortens training time to 5983.54 s. Particularly noteworthy is ResdenseNet’s superiority over other models in terms of having the fewest parameters, the shortest training time, and the highest accuracy, especially when dealing with wheat, barley, and maize datasets. © The Author(s), under exclusive licence to Springer-Verlag London Ltd., part of Springer Nature 2025.