AITP-YOLO: improved tomato ripeness detection model based on multiple strategies

Abstract

Introduction: This paper offers a multi-scale AITP-YOLO model, based on the enhanced YOLOv10s model, to address the challenges of difficult identification and frequent misdetection of tomatoes, facilitating ripeness detection under realistic conditions. Methods: A four-head detector incorporates a small target detection layer, enhancing the model's capacity to identify small targets. Secondly, a multi-scale feature fusion technique employing cross-level features is implemented in the feature fusion layer to amalgamate convolutions of varying sizes, enhancing the model's fusion capacity and generalization proficiency for features of diverse scales. The bounding box loss function is modified to Shape-IoU, with the loss computed by emphasizing the shape and scale of the bounding box, hence enhancing the precision of bounding box regression, expediting model convergence, and augmenting model correctness. Ultimately, the model is compressed via Network Slimming puring,which removes redundant channels while mataining detection accuracy. Results: The experimental findings indicate that the enhanced model achieves average precision, accuracy, and recall of 92.6%, 89.7%, and 87.4%, respectively. In comparison to the baseline network YOLOv10s, the model weights are compressed by 7.64%, while average precision, accuracy, and recall are elevated by 4.6%, 5.8%, and 7.3%, respectively. Discussion: The enhanced model features a reduced model size while exhibiting superior detection capabilities, enabling more efficient and precise recognition of tomato stages amidst complicated backgrounds, hence offering a valuable technical reference for automated tomato harvesting technology.