YOLOV8-MR: An Improved Lightweight YOLOv8 Algorithm for Tomato Fruit Detection

Abstract

As one of the widely planted fruit and vegetable crops, real-time monitoring of the tomato growth process is crucial for enhancing production and maintaining quality. To fulfill the requirement for real-time detection, we introduce an enhanced tomato target detection approach utilizing the YOLOv8n model, offering essential technical assistance for both real-time monitoring and automated tomato harvesting. Initially, we introduce Receptive-Field Attention Convolution (RFAConv) in the model’s backbone to replace the original Conv operation, enabling the model to process details and complex patterns in images more effectively. Secondly, to enhance the small target detection ability and feature extraction effect of the model, we design a non-local channel attention mechanism (MHNCA), which directs the model to focus more on important channel information during feature extraction and further enhances feature expression through cross-channel interaction. Finally, the RepNCSPELAN4_low module is used to replace the original Cf2 module in the head layer, which further reduces the model parameters, reduces calculation costs and speeds up the inference speed. To assess the efficacy of our enhanced model, we conducted comparisons with YOLOv7, YOLOv9, YOLOv10, and several other models. Experiments demonstrate that, compared with YOLOv8n, the improved YOLOv8n model shows a 1.7% boost in mAP% value, a 1.6% boost in accuracy, a reduction in model parameters from 3.01M to 2.83M, and a detection speed of 121.13 FPS, which meets the requirements for real-time detection. In comparison to other models, the improved model has enhanced overall performance, providing a theoretical foundation for the tomato detection system and enabling efficient and high quality harvesting.