Automatic Correction of Labeling Errors Applied to Tomato Detection

Abstract

Accurate labeling is critical for training reliable deep learning models in agricultural applications. However, manual labeling is often error-prone, especially when performed by non-experts, and such errors (modeled as noise) can significantly degrade model performance. This study addresses the problem of correcting labeling errors in object detection datasets without human intervention. We hypothesize that label noise can be reduced by exploiting the feature space representation of the data, enabling automatic refinement through repeated model-based filtering. To test this, we propose a recursive methodology that employs a YOLOv5 detector to iteratively relabel a dataset of Prunaxx and Paipai tomato images captured in greenhouse environments. The correction process involves training the detector, predicting new labels, and replacing existing labelings over multiple iterations. Experimental results show substantial improvements: the mean Average Precision at an IoU threshold of 0.50 (mAP-50) increased from 0.8 to 0.86, the mean Average Precision across IoU thresholds from 0.50 to 0.95 (mAP-50:95) increased from 0.46 to 0.63, and Recall improved from 0.68 to 0.82. These results demonstrate that the model was able to detect more true positives after filtering, while also achieving more accurate bounding box predictions. Although a slight decrease in Precision was observed in later iterations due to false positives, the overall quality of the dataset improved consistently. In conclusion, the proposed filtering method effectively enhances label quality without manual intervention and offers a scalable solution for improving object detection datasets in precision agriculture.

Abstract from the paper’s OpenAlex record; open copy licensed cc-by. Read at the source ↗

The text the model read, with the source record’s inline markup (italics, sub- and superscripts) rendered and its character entities decoded; quote offsets refer to the raw text.