Ta-YOLO: overcoming target blocked challenges in greenhouse tomato detection and counting

Screening and cultivating healthy small tomatoes, along with accurately predicting their yields, are crucial for sustaining the economy of tomato industry. However, in field scenarios, counting small tomato fruits is often hindered by environmental factors such as leaf shading. To address this challenge, this study proposed the Ta-YOLO modeling framework, aimed at improving the efficiency and accuracy of small tomato fruit detection. We captured images of small tomatoes at various stages of ripeness in real-world settings and compiled them into datasets for training and testing the model. First, we utilized the Space-to-Depth module to efficiently leverage the implicit features of the images while ensuring a lightweight operation of the backbone network. Next, we developed a novel pyramid pooling module(DASPPF) to capture global information through average pooling, effectively reducing the impact of edge and background noise on detection. We also introduced an additional tiny target detection head alongside the original detection head, enabling multi-scale detection of small tomatoes. To further enhance the model's focus on relevant information and improve its ability to recognize small targets, we designed a multi-dimensional attention structure(CSAM) that generated feature maps with more valuable information. Finally, we proposed the EWDIoU bounding box loss function, which leveraged a 2D Gaussian distribution to enhance the model's accuracy and robustness. The experimental results showed that the number of parameters, FLOPs, and FPS of our designed Ta-YOLO were 10.58M, 14.4G, and 131.58, respectively, and its mean average precision(mAP) reached 84.4%. It can better realize the counting of tomatoes with different maturity levels, which helps to improve the efficiency of the small tomato production and planting process.