A transformer-enhanced real-time detection method for young fruits in complex orchard environments

Abstract

Accurate detection of apple young fruits is a critical prerequisite for achieving automated orchard management. However,due to the small size of young fruits, their similarity in color to the background, and frequent overlap and occlusion, real-time detection in complex orchard environments becomes a significant challenge. To overcome these challenges, we propose a multi-level, end-to-end, Transformer-enhanced real-time detection method–HLEdetection. This architecture is composed of three core modules: (1) a Hierarchical Multi-scale Feature Extraction module, which extracts fruit features at multiple resolutions to enhance the perception of small objects; (2) a Learned Positional Encoding Injection module, which introduces learnable positional embeddings during the decoding stage to strengthen feature alignment across scales, thereby improving localization precision and semantic consistency; and (3) the EMASlideLoss function, which combines an exponential moving average strategy with IoU-based weighting to mitigate severe class imbalance, further enhancing training stability and accuracy. We construct a custom dataset comprising four representative orchard scenarios to rigorously evaluate the proposed method. Experimental results demonstrate that HLEdetection achieves an average Precision of 91.7% and Recall of 79.7% across complex scenes, while maintaining a real-time inference speed of 27.8 FPS ( ≈ 36.0 ms/frame), thereby achieving an optimal balance between accuracy and efficiency. Compared with mainstream object detectors such as the YOLO series, DynamicDet, Faster R-CNN, and RT-DETR, our model shows over 20% improvement in both F1 score and mAP50-95, particularly excelling in densely occluded environments. These results confirm that the proposed method offers a robust and efficient solution for young fruit detection, providing solid technical support for the development of intelligent orchard equipment such as robotic harvesting, precision thinning, and targeted spraying systems.The source code is available at https://github.com/lijinyan628-ops/-recognition-of-apple .

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.