Abstract:

In view of the current mainstream deep neural network model, which aims to improve the accuracy and ignores the real-time performance and the size of the model, a lightweight target recognition deep neural network is proposed. Based on the high-efficiency convolution method such as depth separation convolution and packet convolution, an invariant resolution convolution module and a downsampling module for image feature extraction are designed. Based on this, a deep backbone network is constructed and the branch is reduced. The target recognition model of visual perception is experimentally verified on the dataset, and 72.7% of mAPs are obtained. The inference speed reaches 66.7 frames per second on the NVIDIA 1080Ti GPU.

Key words: target recognition, lightweight, pruning

摘要：

针对当前一些主流的深度神经网络模型旨在追求准确率的提升，而忽略了模型的实时性及模型大小问题，提出了一种轻量级目标识别深度神经网络。基于深度分离卷积、分组卷积等轻量化的高效卷积方式，设计了用于图像特征提取的不变分辨率卷积模块和下采样模块，并依此构建了深度主干网络，并对网络进行了减枝。在创建的数据集上对视觉感知的目标识别模型进行了实验验证，获得了72.7%的mAP，在NVIDIA 1080Ti GPU上推理速度达到66.7 帧/s。

关键词: 目标识别, 轻量化, 剪枝