A Lightweight Object Detection Network Based on Dense Dilated Encoder Net

• 1、Beijing University of Posts and Telecommunications

Abstract：Aiming at the difficulty of running the CenterNet network with a large number of parameters and computation on devices with limited computing power and resources, a lightweight object detection network based on dense dilated encoder is proposed. Referring to the ShuffleNetv2 network based on depthwise separable convolutions, the backbone of CenterNet is optimized. A Light Dense Dilated Encoder (LDDE) is constructed by introducing multiple concatenated dilated convolutions and performing dense skip-layer connections. LDDE expands the receptive field at multiple scales by changing the dilation rate of dilated convolutions. which significantly improves the object detection performance. On the PASCAL VOC dataset, the mean average precision of LDDE-Net iA Lightweight Object Detection Network Based on Dense Dilated Encoder Nets higher than the lightweight model YOLOv4-tiny, and the parameters are only 68.65% of YOLOv4-tiny. On the NVIDIA GeForce RTX 2080 Ti, LDDE-Net has an FPS of 106.1. On the embedded platform NVIDIA Jetson Nano, the inference time of LDDE-Net is 120ms.

Keywords： pattern recognition and intelligent systems dilated convolution lightweight, computer vision deep learning