.png){kind=icon}
.png){kind=icon}
Return
Multi-scale convolution and dynamic task interaction detection head for efficient lightweight plum detection
.png){kind=icon}
PRE
AI
DOI：10.1016/j.fbp.2024.12.007.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Automating fruit detection is crucial for boosting efficiency and ensuring high-quality produce. Plums, scientifically known as Prunus salicina, presents a unique challenge due to their soft texture, vulnerability to damage, and frequent instances of overlapping. These traits complicate manual inspection and present significant obstacles for automated systems. In addition, plum detection devices often face limitations in computing power and storage space, demanding high performance in terms of model lightweighting. To meet the requirements for detection performance and lightweighting in practical plum detection, we propose a novel, high-precision, lightweight plum detection model based on YOLOv8n, named Efficient Lightweight Plum Detector (ELPD). Our model introduces several innovations: we propose Potently Efficient Multi-Scale Convolution (PEMSConv) to enhance the model's capability in extracting multi-scale features, improving detection accuracy while reducing model size. Additionally, we introduce Dynamic Task Interaction Detection Head (DTIDH) to enhance the interaction between classification and localization tasks, boosting detection performance while the use of shared convolution further reduces the model size and parameters. Moreover, we propose Focaler-Minimum Point Distance Intersection over Union (Focaler-MPDIoU), which enables the model to focus on the majority of samples in the dataset based on their difficulty, further enhancing detection performance. Focaler-MPDIoU also considers various geometric properties of bounding boxes, accelerating model convergence. Finally, we evaluate our model using a dataset from a plum orchard. In terms of detection performance, our model outperforms the baseline model with improvements of 2.03% in mAP@0.5 and 1.02% in mAP@0.5:0.95. Regarding lightweight efficiency, our model achieves reductions of 32.53% in model size and 35.19% in parameters compared to the baseline.
Keywords:
Plum
Computer vision
Lightweight network
Journal
IF:
3.4
Papers:
2.7K
Citations:
7.1K