A Lightweight Pest Detection Model for Drones Based on Transformer and Super-Resolution Sampling Techniques

Abstract

With the widespread application of drone technology, the demand for pest detection and identification from low-resolution and noisy images captured with drones has been steadily increasing. In this study, a lightweight pest identification model based on Transformer and super-resolution sampling techniques is introduced, aiming to enhance identification accuracy under challenging conditions. The Transformer model was found to effectively capture spatial dependencies in images, while the super-resolution sampling technique was employed to restore image details for subsequent identification processes. The experimental results demonstrated that this approach exhibited significant advantages across various pest image datasets, achieving Precision, Recall, mAP, and FPS scores of 0.97, 0.95, 0.95, and 57, respectively. Especially in the presence of low resolution and noise, this method was capable of performing pest identification with high accuracy. Furthermore, an adaptive optimizer was incorporated to enhance model convergence and performance. Overall, this study offers an efficient and accurate method for pest detection and identification in practical applications, holding significant practical value.

Few-Shot Image Classification of Crop Diseases Based on Vision-Language Models

“…Kaya et al [11] investigated and substantiated the utility of transfer learning models in facilitating crop classification. Bai et al [12] introduced a rice disease detection method grounded in multi-source data and transfer learning, significantly boosting disease recognition accuracy. Transfer learning eases the demand for extensive training samples to some extent, yet it does not resolve the overfitting issue that arises when models are trained with limited data.…”