基于YOLOv8的规则形状检测与自动标注系统设计

Design of Regular Shape Detection and Automatic Annotation System Based on YOLOv8

摘要:

针对受控可见光场景下规则形状目标精准识别与高效标注的实际需求，研究构建轻量化的目标检测与自动标注集成框架，采用YOLOv8作为基础检测模型，融合卷积块注意力模块（CBAM, Convolutional Block Attention Module）优化特征提取过程，基于自建的16类单色背景规则形状数据集开展模型训练与验证；设计从检测结果到YOLO格式标注文件的端到端转换流程，实现规则形状目标的自动化检测与标注；通过INT8量化技术对模型进行轻量化处理，结合RK1828 AI协处理器完成边缘端部署，并开展多硬件平台的推理性能测试。经3次独立验证，该框架在训练集与验证集8:2划分的数据集上mAP50达到0.992±0.3%，在NVIDIA RTX 3060显卡上单帧推理耗时约6.9 ms；模型量化至5.1 MB后，在RK1828上的推理延迟为12.8 ms，mAP50仅下降0.5%至0.987；使用自动标注文件训练的模型mAP50达0.976，与人工标注模型的0.981相差仅0.5%，标注质量满足模型再训练需求。该集成框架在规则形状检测与自动化标注场景中具备高精度、高实时性的应用优势，可适配桌面端与边缘端多平台部署，但其在复杂工业和安防场景中的泛化性能仍需进一步验证优化。

Abstract:

Aiming at the practical demand for accurate recognition and efficient annotation of regular shape targets in controlled visible light scenes, a lightweight integrated framework for object detection and automatic annotation is researched and constructed. YOLOv8 is adopted as the basic detection model, and the Convolutional Block Attention Module (CBAM) is fused to optimize the feature extraction process. Model training and verification are carried out based on a self-constructed 16-category regular shape dataset with solid-color background. An end-to-end conversion process from detection results to YOLO format annotation files is designed to realize automatic detection and annotation of regular shape targets. The model is lightweight by INT8 quantization technology, deployed on the edge side combined with RK1828 AI coprocessor, and the inference performance tests on multiple hardware platforms are carried out. After 3 independent verifications, the mAP50 of the framework reaches 0.992±0.3% on the dataset with an 8:2 division of training and validation sets, and the single-frame inference time is about 6.9 ms on the NVIDIA RTX 3060 graphics card. After the model is quantized to 5.1 MB, the inference delay on RK1828 is 12.8 ms, and the mAP50 only drops by 0.5% to 0.987. The mAP50 of the model trained with automatically annotated files reaches 0.976, which is only 0.5% different from 0.981 of the model trained with manual annotation, and the annotation quality meets the requirements of model retraining. The integrated framework has the application advantages of high precision and high real-time performance in the scenarios of regular shape detection and automatic annotation, and can be adapted to multi-platform deployment on desktop and edge sides, but its generalization performance in complex industrial and security scenarios still needs further verification and optimization.