An improved YOLOv7 model based on Swin Transformer and Trident Pyramid Networks for accurate tomato detection

Abstract

Accurate fruit detection is crucial for automated fruit picking. However, real-world scenarios, influenced by complex environmental factors such as illumination variations, occlusion, and overlap, pose significant challenges to accurate fruit detection. These challenges subsequently impact the commercialization of fruit harvesting robots. A tomato detection model named YOLO-SwinTF, based on YOLOv7, is proposed to address these challenges. Integrating Swin Transformer (ST) blocks into the backbone network enables the model to capture global information by modeling long-range visual dependencies. Trident Pyramid Networks (TPN) are introduced to overcome the limitations of PANet’s focus on communication-based processing. TPN incorporates multiple self-processing (SP) modules within existing top-down and bottom-up architectures, allowing feature maps to generate new findings for communication. In addition, Focaler-IoU is introduced to reconstruct the original intersection-over-union (IoU) loss to allow the loss function to adjust its focus based on the distribution of difficult and easy samples. The proposed model is evaluated on a tomato dataset, and the experimental results demonstrated that the proposed model’s detection recall, precision, F 1 score, and AP reach 96.27%, 96.17%, 96.22%, and 98.67%, respectively. These represent improvements of 1.64%, 0.92%, 1.28%, and 0.88% compared to the original YOLOv7 model. When compared to other state-of-the-art detection methods, this approach achieves superior performance in terms of accuracy while maintaining comparable detection speed. In addition, the proposed model exhibits strong robustness under various lighting and occlusion conditions, demonstrating its significant potential in tomato detection.