欢迎访问《图学学报》 分享到：
图学学报

图学学报 ›› 2021, Vol. 42 ›› Issue (2): 198-205.DOI: 10.11996/JG.j.2095-302X.2021020198

• 图像处理与计算机视觉 • 上一篇    下一篇

基于模型压缩的 YOLOV3 实时枪支识别方法

  

  1. 1. 中国人民武装警察部队工程大学信息工程学院，陕西 西安 710086；  2. 中国人民武装警察部队工程大学反恐指挥信息工程研究团队，陕西 西安 710086；  3. 中国人民武装警察部队工程大学研究生大队，陕西 西安 710086
  • 出版日期:2021-04-30 发布日期:2021-04-30
  • 基金资助:
    国家自然科学基金项目(U1603261)；新疆维吾尔自治区自然科学基金资助项目(2016D01A080) 

Real-time gun detection method based on compressed YOLOV3 

  1. 1. School of Information Engineering, Engineering University of PAP, Xi’an Shaanxi 710086, China;  2. Counter-Terrorism Command Information Engineering Research Team, Engineering University of PAP, Xi’an Shaanxi 710086, China;  3. Graduate Team, Engineering University of PAP, Xi’an Shaanxi 710086, China
  • Online:2021-04-30 Published:2021-04-30
  • Supported by:
    National Natural Science Foundation of China (U1603261); Natural Science Foundation of Xinjiang Uygur Autonomous Region (2016D01A080) 

PDF (PC)

121

摘要： 以枪支为代表的危险目标检测一直是安防领域的重要研究之一。针对当前人工通过监控视频检 查枪支等危险物效率低且准确率易受检查人员工作时长影响的问题，提出了利用剪枝方法对 YOLOV3 模型做 压缩的实时枪支检测方法。采用 K-means ++算法对图像样本进行锚定框 Anchor 大小聚类，以提高模型精度。 利用“通道+层”剪枝方法将训练后的模型进行压缩，通过模型修正恢复压缩前的精度。实验结果表明，该方 法在保持较高精度的情况下，不仅降低了模型对内存资源的占用，且进一步减少计算量，大大提高了模型推理 速度。与 YOLOV3 方法相比，该方法在 jetson nano 平台上对模型参数的缩减比例达到 1/52，推理速度提高了 6 倍，而精确度几乎保持不变，从而达到对枪支危险物检测的实时性和高精度要求。

关键词: 枪支检测, YOLOV3, 模型压缩, K-means++, 实时检测

Abstract: The detection of such dangerous targets as guns has always been one of the important research subjects in the field of security. Manual inspection of guns and other dangerous objects through surveillance video is inefficient, the accuracy of which can be easily affected by unfavorable conditions of inspectors due to continuous working. Therefore, a real-time gun detection method using the pruning method to compress the YOLOV3 model was proposed. In order to improve model accuracy, the K-means ++ algorithm was adopted to cluster Anchor size of data set. Then the pruning method of “channel + layer” was employed to compress the trained model. Finally, by means of retraining, the accuracy was restored to that before compression. The experimental results show that while still maintaining high accuracy, this method can reduces not only the model’s occupation of memory resources, but also the computing load, greatly improving the inference speed of the model. Compared with the YOLOV3 method, this method reduces the model parameters by up to 1/52 on the jetson nano platform and increases the reasoning speed by 6 times, with the accuracy almost unchanged, thus meeting the requirements of real-time and high-precision detection of such dangerous objects as guns. 

Key words: gun detection, YOLOV3, model compression, K-means ++, real-time detection 

中图分类号: