YOLOv3- Tiny's Weight Size Reduction using Pruning and Quantization

CNN has been widely applied to various technology fields, like real-time edge devices. One of the architecture detection objects based on CNN is YOLO, a real-time detection object with a high level of accuracy. Detection objects on the edge device itself are expected to have speed, accuracy, and less memory needed to carry out the detection process. In this paper, YOLOv3- Tiny will be pruned and quantized to reduce the size of the data from the weight without significant reduction of accuracy. Experimental results show that YOLOv3- Tiny weight data decreased up to 85.7% compared to the weight data before the pruning process, while in the quantization mode. Moreover, the iterative pruning method can reduce the weight up to 43%.