A compression pipeline for one-stage object detection model

摘要: Deep neural networks (DNNs) have strong fitting ability on a variety of computer vision tasks, but they also require intensive computing power and large storage space, which are not always available in portable smart devices. Although lot studies contributed to the compression image classification networks, there few model algorithms for object detection models. In this paper, we propose general pipeline one-stage meet real-time requirements. Firstly, softer pruning strategy backbone reduce number filters. Compared with original direct pruning, our method can maintain integrity network structure drop accuracy. Secondly, transfer knowledge small by distillation accuracy caused pruning. Finally, as edge devices more suitable integer operations, further transform 32-bit floating point into 8-bit through quantization. With pipeline, size inference time compressed 10% or less original, while mAP is only reduced 2.5% less. We verified that performance Pascal VOC dataset.