Comparative Evaluation of CNN and Transformer-Based Instance Segmentation Architectures for Whitefly Detection in Citrus Crops

This study focuses on the detection of pests in citrus fruits using computer vision (CV) techniques and image detection technologies based on convolutional neural networks (CNNs). Early detection enables pest control, reduces the use of pesticides, and generates alerts for the inspection of surrounding areas. One of the models used is You Only Look Once (YOLO), which is based on a single-stage CNN. This model is fast, resource-efficient, and has low latency, enabling it to run on a personal computer. Another type of neural network is based on the vision transformer, a paradigm that requires more hardware resources than YOLO but offers greater precision. Therefore, a comparison is carried out between CNN models to evaluate their precision and demonstrate that YOLO is an excellent option compared to vision transformer-based models, which are more resource-intensive. Through the capture of images in the field, it will be shown that the difference in precision compared to other models that are more demanding in terms of hardware is minimal. The study also reveals that all the models assessed have difficulties in detecting small objects, suggesting a need to improve this through software techniques or hardware enhancements, such as the optics of drone lenses. A wide-angle lens could be used for preliminary detection, and another with a higher zoom for confirmation and the detection of small insects, thereby facilitating its adoption where computational resources are limited.