Abstract: To rapidly and accurately detect the maturity levels of Ding’ao bayberry (Myrica rubra) in complex natural growth environments, an improved Faster R-CNN model (ConvNeXt-T+SE+FPN)-based maturity detection method was proposed. ConvNeXt-T was adopted as the backbone feature extraction network to enhance detection capabilities in complex scenarios. The SE attention mechanism and Feature Pyramid Network (FPN) were introduced to improve the model’s sensitivity to maturity-related features and detection of small-target fruits. Compared to ResNet50, ConvNeXt-T+SE, ConvNeXt-T+FPN, and ConvNeXt-T+SE+FPN increased the mean average precision (mAP) by 14.75%, 19.85%, and 21.86%, respectively. The ConvNeXt-T+SE+FPN configuration achieved the largest mAP improvement, effectively enhancing detection performance for different maturity levels of Ding’ao bayberry. Through training and testing on the Ding’ao bayberry image dataset, the improved Faster R-CNN model demonstrated high accuracy in detecting different maturity levels. The average precision (AP) for unripe, semi-ripe, near-ripe, and fully ripe fruit recognition was 96.90%, 94.63%, 95.91%, and 97.58%, respectively, with an mAP of 96.26%. Compared to the original Faster R-CNN model, the improved model achieved a 21.86% increase in mAP. The improved Faster R-CNN model effectively enhanced the detection accuracy of Ding’ao bayberry maturity, providing strong support for intelligent harvesting of bayberry fruits.

Key words: Ding’ao bayberry (Myrica rubra), improved Faster R-CNN model, maturity, detection