DMP-YOLO: Dense multi-scale perception for complex scenes YOLO algorithm Prunus humilis small target detection

• To address the problem of small Prunus humilis target loss in complex backgrounds, we designed the C2f_MSCA (Multi-Scale Context Aggregation Block) module for extracting information from small Prunus humilis fruit targets. This module enhances the feature extraction capability for small Prunus humilis targets, effectively suppresses background noise interference, and improves small target extraction efficiency. • To resolve target loss caused by the dense arrangement of Prunus humilis fruits, SPP and TransformerBlock modules are introduced into the last two layers of the backbone network to establish long-range dependencies, thereby improving target extraction efficiency. • To tackle target loss issues arising from overlapping areas between branches, leaves, and fruits, we propose the "Multi-Dimensional Adaptive Convolution" C2f_MDAC (Multi-Dimensional Adaptive Convolution). This module adaptively fuses target information across different scales to enhance the fusion efficiency of multi-scale Prunus humilis fruits. • This study provides theoretical basis and technical guidance for the intelligent production of Prunus humilis (Chinese dwarf cherry) fruits. To address the critical challenges of low detection accuracy and target loss in multi-scale Prunus humilis small fruit detection under severe occlusion by branches and leaves, this study proposes an improved DMP-YOLO model based on YOLO-v8. First, a multi-channel attention mechanism (C2f_MSCA module) is integrated into the deep layers of the YOLO-v8 backbone network. This module enhances multi-scale fruit feature extraction capabilities, suppresses background noise interference, and improves small target recognition. Second, Spatial Pyramid Pooling (SPP) and TransformerBlock modules replace the original C2f modules in the final two backbone layers, establishing long-range dependencies and global contextual relationships to mitigate target loss. Third, a "Multi-Dimensional Adaptive Convolution" module (C2f_MDAC) is embedded into the bottleneck, adaptively fusing multi-scale target information to optimize the detection efficiency of Prunus humilis fruits across varying sizes. Extensive experiments demonstrate that the improved model effectively enhances detection accuracy for small Prunus humilis targets in complex backgrounds with severe occlusions, both under normal lighting and dim conditions. Under normal lighting, the DMP-YOLO model achieves an average detection accuracy of 73.5 %, outperforming YOLO-v3. YOLO-v5, YOLO-v6, and YOLO-v8 by 4.7, 23, 19.3, and 14.3 percentage points respectively. Under dim lighting, the DMP-YOLO model achieves an average detection accuracy of 69.8 %, surpassing YOLO-v8, YOLO-v10, YOLO-v11, and YOLO-v12 by 12.9, 12.9, 16.9, and 16.5 percentage points respectively. Under normal lighting conditions, the DMP-YOLO model achieves an inference speed of 0.9 milliseconds per image, compared to 18.8, 10.9, 6.1, 2.1, 8.7 ms, 9.2 ms, and 9.0 ms respectively. Under dim lighting conditions, the detection inference speed is 7.1 ms, outperforming the other four models by 3.4 ms, 3.6 ms, 7.2 ms, and 5.5 ms respectively, thereby meeting the real-time detection requirements for precision agriculture. By conducting precise monitoring at the early stage of fruit development (especially the very small fruit stage), and based on the observed different morphological traits, this approach predicts the fruit set rate, maturity, and quality of Prunus humilis fruits. It provides a scientific basis and intelligent decision-making support for formulating precise and differentiated field water and fertilizer management strategies.