RetinaNet and Vision Transformer-Based Model for Wheat Head Detection

Vision transformers have achieved cutting-edge results on numerous object detection benchmarks, showcasing its potency as a robust object detection framework. Our aim is to design and develop a vision transformer based deep learning model towards the goal of smart agriculture. Wheat head detection is an important task in precision agriculture for estimating crop yields and monitoring plant health. This research study proposes a two-stage detector system by combining the RetinaNet object detection architecture with the vision transformer to improve the wheat head detection performance. RetinaNet is used as a proposal generator to generate candidate bounding boxes, and ViT is used as a backbone network for object classification/localization. The model was evaluated using Focal loss and Smooth L1 loss functions to jointly optimize classification and bounding box regression performance. The combination of RetinaNet and ViT took advantage of the strengths of both the approaches. RetinaNet generated candidate bounding boxes with high recall, while ViT processed these candidates efficiently and accurately, potentially reducing the number of false positives and improving overall detection accuracy. Experimental results on a wheat head detection dataset demonstrated that the proposed RetinaNet+ViT model is a promising and potentially efficient approach for wheat head detection, and has shown promising results in object detection task.