An Improved YOLOv3-based Neural Network for De-identification Technology
Ji-Hun Won, Dong‐Hyun Lee, Kyungmin Lee, Chi-Ho Lin · 2019 34th International Technical Conference on Circuits/Systems, Computers and Communications (ITC-CSCC) · 2019
In this paper, we proposed an improved YOLOv3-based neural network for De-identification technology. The existing YOLOv3 is a network with fast speed and performance recently. Most surveillance system using CCD cameras simultaneously store images from cameras installed in multiple locations. In such an environment, the use of deep learning requires a method of detecting objects through a single inference engine in a plurality of image. If the inference engine hardware is used for each camera channel, the cost of building a surveillance system increases significantly. Therefore, in the field of surveillance systems, a network structure with a high detection speed is required even if the detection performance is slightly degraded. This paper proposes a method to increase the detection speed by reducing the existing YOLOv3 network Architecture. 53 feature extractors, Darknet-53, are reduced to 24 layers. Therefore, a total of 106 layers is reduced to 39. And 53 YOLOv3 box detection parts are reduced to 15 layers. In order to verify the efficiency of the proposed algorithm, the WIDER FACE dataset and its own collected dataset, we compared the performance with the existing YOLOv3-tiny and YOLOv3. As a result, the result was 87.48% mAP improved by 19.55% compared to the conventional YOLOv3-tiny. And I got a slow result of 100.5 FPS speed than the existing speed. And Object De-identification technology has been applied according to the results of the detection box. Therefore, it is faster than YOLOv3, and it is similar to YOLOv3 detection accuracy, proving that it is better than YOLOv3-tiny in real time detection.