Liu Fei, Hu Mingyue
Shanghai Polytechnic University, Shanghai, China.
PeerJ Comput Sci. 2025 Apr 15;11:e2832. doi: 10.7717/peerj-cs.2832. eCollection 2025.
The advancement of assistive technologies for individuals with disabilities has increased the demand for efficient and accurate object detection algorithms, particularly in meal-assistance robots designed to identify and handle food items such as fruit chunks. However, existing algorithms for fruit chunk detection often suffer from prolonged inference times and insufficient accuracy.
We propose an improved YOLOv8n algorithm optimized for real-time, high-accuracy fruit chunk detection. The Universal Inverted Bottleneck (UIB) module has been integrated into the original C2f structure, significantly reducing the model's parameter count while preserving detection accuracy. Furthermore, the coordinate attention (CA) mechanism has been incorporated into the detection head to enhance the focus on fruit chunk regions within complex backgrounds while suppressing irrelevant features, thus improving detection performance. Additionally, the ADown module from YOLOv9 has been embedded into the YOLOv8 backbone network, further increasing accuracy and reducing the number of parameters.
Experimental results indicate that these enhancements substantially improve detection accuracy while reducing model size. Specifically, the optimized model achieves a 1.9 MB reduction in size, a decrease of 2.5 GFLOPs in parameter count, and an increase in mAP50 and mAP50-95 by 2.1% and 3.3%, respectively. The improved algorithm (UCA-YOLOv8n) enables real-time, accurate detection of various fruit chunks. Comparative analyses with other mainstream object detection algorithms further demonstrate the superiority and effectiveness of the proposed method.
辅助技术的进步增加了对高效准确的目标检测算法的需求，尤其是在旨在识别和处理诸如水果块等食品的用餐辅助机器人中。然而，现有的水果块检测算法往往推理时间长且准确性不足。
我们提出了一种针对实时、高精度水果块检测进行优化的改进型YOLOv8n算法。通用倒置瓶颈（UIB）模块已集成到原始的C2f结构中，在保持检测精度的同时显著减少了模型的参数数量。此外，坐标注意力（CA）机制已被纳入检测头，以增强在复杂背景下对水果块区域的关注，同时抑制无关特征，从而提高检测性能。此外，YOLOv9中的ADown模块已嵌入到YOLOv8骨干网络中，进一步提高了准确性并减少了参数数量。
实验结果表明，这些改进在减少模型大小的同时大幅提高了检测精度。具体而言，优化后的模型大小减少了1.9MB，参数数量减少了2.5 GFLOPs，mAP50和mAP50-95分别提高了2.1%和3.3%。改进后的算法（UCA-YOLOv8n）能够实时、准确地检测各种水果块。与其他主流目标检测算法的对比分析进一步证明了所提方法的优越性和有效性。
