Comparative Evaluation of Convolutional and TransformerBased Models for Cucumber Disease Classification

Abstract

Cucumber (Cucumis sativus) is an important horticultural crop, and disease-related losses can reduce both yield and market quality. Image-based deep learning methods have been increasingly used for plant disease recognition, yet practical model selection requires attention to both predictive performance and computational cost. In this study, four pretrained deep learning architectures were evaluated for eight-class cucumber disease classification: DenseNet201, Inception v4, ConvNeXt v2 Base, and DeiT Base. The experiments were conducted on a public dataset containing 7,689 images, including two healthy and six diseased categories, using a 70% training, 15% validation, and 15% test split. DeiT Base achieved the highest test accuracy among the evaluated models, with an accuracy of 0.9957. Inception v4 followed closely with an accuracy of 0.9948, whereas DenseNet201 provided a lower but still competitive accuracy of 0.9846 with the smallest computational burden, requiring 18.11M parameters and 8.58 GFLOPs. The results indicate that highly accurate cucumber disease classification can be achieved with both convolutional and Transformer-based models, but the preferred architecture depends on whether the priority is maximum accuracy or lower computational cost. Since the experiments were limited to a single public dataset, further validation on fieldacquired images is needed before drawing conclusions about real-world deployment.