Detecting the maturity of Lonicera caerulea berries in the field using improved YOLOv11s

Blue honeysuckle is one of the small fruit crops with high nutritional properties. Significant challenges remain to extract the image feature during field growth, due to the clustering of fruits at different maturity levels, partial occlusion by branches and leaves, as well as highly variable illumination caused by weather and sun angles. Previously, the high mean average precision has been achieved in the single-stage object detection, particularly improved YOLO models with the attention mechanisms, optimal networks of feature extraction, or small-target detection. Strong performance and applicability can also be found to detect the fruit maturity in complex field environments. The YOLO framework also shared robust generalization. However, the existing YOLO models are often highly specific to the growth environment, planting patterns, such as the clustered and dense fruiting, and detection metrics of fruits or vegetables. This specificity has caused the low transferability over different objects, metrics, and requirements. It is often required for the real-time deployment on resource-constrained mobile devices in the field. This study aims to improve the accuracy and precision of maturity identification on blue honeysuckle fruits in field environments. A YOLOv11s-ACM model was also proposed using the YOLOv11s algorithm. 1) The Attention-based Intrascale Feature Interaction module was introduced into the backbone network. The high accuracy was obtained in detecting the blue honeysuckle maturity under field conditions. A globally content-adaptive attention mechanism was incorporated to selectively emphasize the informative features within the same scale. High accuracy was enhanced to detect the blue honeysuckle maturity in the field. 2) C3K2 convolution in the backbone network was replaced with a C3K2 structure with Dynamic Snake Convolution. Residual connections and dynamic feature aggregation were used to adaptively capture elongated structures. The adaptability of the model was improved in the occluded and low-illumination scenes after modification. Thereby, the better performance was enhanced to detect the irregularly shaped fruits of blue honeysuckle in field environments. 3) A Multi-Separated and Enhancement Attention Module Head was introduced to strengthen the dense targets with multi-scale features under field conditions. Experimental results demonstrate that the YOLOv11s-ACM achieved a 4.5 percentage point improvement in mean average precision and a 27% reduction in the inference time per image, compared with the baseline YOLOv11s model. Furthermore, compared with the YOLOv5s, YOLOv7-Tiny,YOLOv7,YOLOv8s, YOLOv11n, YOLOv11s, and YOLOv11m, the YOLOv11s-ACM achieved the mAP improvements of 6.0, 5.9, 2.6, 5.8, 12.8, 4.5, and 5.5 percentage points, respectively, indicating significantly enhanced overall performance. Mobile deployment validated that the high accuracy and inference speed fully met the requirements for the rapid and accurate detection of blue honeysuckle fruits under various field conditions, including different lighting scenarios, such as direct sunlight and deep shadows, as well as the occlusion levels ranging from light leaf coverage to heavy branch obstruction. Large-scale real-world datasets can be constructed to further optimize the model adaptability under extreme lighting, severe occlusion, and diverse backgrounds. Additionally, the adaptation and optimization of next-generation YOLO models for mobile terminals can be expected to enhance the model robustness and lightweight deployment in complex environments, ultimately promoting the wide application in modern agriculture.