Pipeline to detect Colorado Potato Beetles as tiny objects under field conditions in real-time using deep learning and transfer learning techniques

Imran Hassan, Ahmad Al-Mallahi, Alimohammad Shirzadifar,
Pipeline to detect Colorado Potato Beetles as tiny objects under field conditions in real-time using deep learning and transfer learning techniques,
Smart Agricultural Technology,
Volume 12,
2025,
101637,
ISSN 2772-3755,
https://doi.org/10.1016/j.atech.2025.101637.
(https://www.sciencedirect.com/science/article/pii/S2772375525008688)
Abstract: This study introduced a novel methodology for real-time identifying Colorado Potato Beetle (CPB), as tiny objects within potato field using cameras mounted on moving sprayer. The methodology consisted of three key steps: targeted image-cropping-based preprocessing, a two-phase transfer learning strategy, and developing an end-to-end detection pipeline system. The image-cropping-based preprocessing step helped preserve the shape and dimension of CPBs during processing by deep learning algorithms, thereby contributing to improved accuracy across the entire detection pipeline. In the two-phase transfer learning approach, the first phase involved training on clear, high-quality focused images, while the second phase fine-tuned the models using field images captured by a camera mounted on a moving sprayer. This approach significantly enhanced the model's robustness to environmental variability, including motion blur. The study compared the performance of three state-of-the-art object detection algorithms, including YOLOv5, YOLOv7, and Faster-RCNN. Thus, the findings from the experiments showed that the image-cropping-based preprocessing approach, when employed in conjunction with transfer learning, substantially increased the detection rates for all studied models. Among all models YOLOv5 emerged as the best choice, offering the best balance between detection accuracy and computational efficiency. At an image size of 640 × 640, YOLOv5 achieved up to 79 % detection accuracy and consistently outperformed others in terms of inference time, processing a single image segment (640 × 640) in just 42 milliseconds enabling real-time performance. Further optimization using Tensor-RT reduced the pipeline’s end-to-end latency to 86 milliseconds per HD image (1920 × 1080), facilitating operation at over 10 frames per second. This research enhances the growing use of deep learning and computer vision in agricultural pest control. The results could be valuable for improving site-specific pesticide applications, benefiting farmers, and protecting the environment.
Keywords: Real-time detection pipeline; Colorado Potato Beetle; Tiny object detection; Transfer Learning; Precision Spraying; YOLO; Faster-RCNN