Comparative Analysis of Transformer-Based and CNN Models for High-Throughput Wheat Head Detection

Wheat head or spike detection is significant for phenotyping because it can be directly correlated to yield and is an indicator of yield potential. Historically, wheat head counting was a labor-intensive and error prone process. Use of Deep Learning (DL) techniques has automated this process allowing automated wheat head detection and counting using high resolution imagery, allowing large-scale, High Throughput Phenotyping (HTP). Despite the use of advanced technologies, wheat head detection is a challenging task due to high environmental variability, cultivar differences, and head overlap. Several attempts have been made to make the DL models more robust and the wheat head datasets more diverse to improve the detection accuracy and reliability. With the introduction of advanced DL architectures, there has been continuous improvement in accuracy of head detection. In this study, we have evaluated the performance of three different cutting-edge DL models - YOLOv10x, RetinaNet, and MM-Grounding DINO for wheat head detection. We have also combined two different wheat datasets, Global Wheat Head Detection (GWHD) 2021 and SPIKE dataset to get a diverse dataset with a wide range of genotypes. This study aims to contribute to the ongoing evolution of wheat head detection techniques and provide an insight into how these three models perform for this task.