An effective citrus ripeness detection model for complex orchard scenarios

To address the challenges faced in detecting citrus ripeness in orchard environment, such as leaf obstructions, overlapping fruits, uneven illumination and abundant small targets, we propose an effective citrus ripeness detection model that integrates dynamic depth-separable convolutions and collaborative attention (DC-YOLO).

(1) It replaces standard convolutions with dynamic depth-separable convolutions to reduce computational complexity and model parameters, while improving feature extraction adaptability. (2) It integrates a collaborative attention mechanism consisting of coordinate attention, cross-scale attention, and dual-feature multi-head self-attention. This module alleviates semantic conflicts during feature fusion and strengthens the representation of multi-scale target features and spatial localization capability. (3) A multi-loss function combining localization, classification, ripeness assessment, and occlusion awareness is adopted to meet the demands of fine-grained classification and occlusion-aware detection for citrus ripeness.

Experimental results demonstrate that for the citrus ripeness detection task, our model achieves a precision of 0.980, a recall of 0.965, an F1-score of 0.972, an mAP50 of 0.975 and an mAP50:95 of 0.769. Compared with the baseline model, its parameters and computational complexity are reduced by 12.79\% and 18.28\%, respectively. The inference FPS reaches 93.42, achieving a satisfactory result between detection accuracy and real-time performance.

Although our method meets the need for real-time detection of citrus ripeness in orchards, there are still challenges regarding the deployment of edge devices and the diversity of citrus data samples. Code is available at \url{https://github.com/zhangq0601/dcyolo}.