| 摘要: |
| 目前防窃照技术手段主要依靠图像采集设备实时采集视频流，对拍照行为进行识别检测，但受复杂环境影响，现有识别方法存在精确度低、实时性差等问题。针对上述问题，提出了一种基于实例语义图的屏幕反窃照识别算法。该算法首先通过图像分割模型U-Net提取原图像的显著视觉目标后生成实例语义图，然后通过微调单阶段的目标检测模型YOLO-v5实现对语义图中感兴趣对象的目标检测，最后通过设计的基于Inception-ResNet的拍照识别二分类模型实现对屏幕窃照行为的识别。实验结果表明，该算法在实际场景下的平均识别准确率达到95%以上。 |
| 关键词: 窃照识别 目标检测 深度学习 语义图 |
| DOI：10.20079/j.issn.1001-893x.220415004 |
|
| 基金项目:中电天奥产业发展基金（202011500402） |
|
| An algorithm for screen anti-stealth recognition based on instance semantic graph |
| TAO Guanhong,FAN Zhenjun |
| (Chengdu Spaceon Group Co.,Ltd,Chengdu 610066,China) |
| Abstract: |
| Existing anti-stealth methods mainly rely on detecting photographing behaviors from video streams captured by cameras.However,in complex environments,existing detecting methods have problems such as low accuracy and poor real-time performance.The authors propose a screen anti-stealth recognition algorithm based on the instance semantic graph.The algorithm first generates a semantic graph with the background information of the input image removed by the image segmentation model U-Net,and then detects the object of interest by fine-tuning the single-stage target detection model YOLO-v5.Finally,the screen-shooting behavior is detected through the designed 2-classification model based on Inception-ResNet.The experimental results show that the average recognition accuracy of the algorithm in the actual scene is above 95%. |
| Key words: shooting recognition object detection deep learning semantic graph |
