CEF-enhanced YOLOv11n: An efficient framework for small object detection of young Redlove apple in orchards

The Redlove apple is characterized by its distinctive red flesh, which extends from the skin to the core, making it particularly well-suited for juicing, brewing, and the extraction of natural red anthocyanin. Detection of young Redlove apple is crucial for assessing thinning effectiveness and predicting yield. Accurate and efficient detection is the primary prerequisite for mechanized and intelligent apple orchard management. However, this task presents significant challenges due to the small size of the targets, frequent occlusions, and complex background conditions. To address these issues, this manuscript proposes a novel detection model based on the You Only Look Once(YOLO) architecture. First, a Heterogeneous Kernel-based convolution (HetConv) module is embedded into the cross stage partial (CSP) block, forming a CSPHet module in the backbone network to enlarge the receptive field and enhance multi-scale feature representation. Second, a 1 × 1 convolutional layer is incorporated into the neck to reduce channel dimensions, thereby compressing the model size and lowering computational costs. Furthermore, an Efficient Multi-scale Attention (EMA) mechanism is introduced to capture pixel-level dependencies through cross-dimensional interactions, strengthening the global feature representation. Additionally, a frequency-adaptive dilated convolution (FADC) module is integrated into the detection head to adaptively adjust dilation rates, improving the model's sensitivity to detailed, high-frequency regions. Finally, the minimum point distance intersection over union (MPDIoU) is adopted to accelerate model convergence by addressing the limitations of the original complete IoU (CIoU) loss when the predicted box and the ground truth box share the same center. The proposed model was evaluated on a self-constructed dataset of young Redlove apple. Ablation studies and comparative experiments involving different attention mechanisms, detection heads, and loss functions were conducted to validate its advantages. Experimental results demonstrate that our model achieves a precision(P) of 90.8%, a recall(R) of 84.8%, and a mean Average Precision (mAP) of 92.4%, outperforming the baseline network by 2.1%, 0.5%, and 1.3%, respectively. These findings indicate that the proposed model performs excellent in detecting young Redlove apple, particularly under conditions of small target size and occlusion, thereby contributing to the development of automated and intelligent management systems for Redlove apple orchards.