深度估计网络的可学习步长轻量化研究

Research of Learned Step Quantization Based Lightweight Depth Prediction Network

目前大量被提出的关于单目视觉深度估计网络研究中其网络结构庞大臃肿,在实际部署中会存在占用大、延迟高的问题.针对以上问题,本文提出了基于可学习步长的量化策略的轻量化深度估计网络.该网络采取特征金字塔(FPN)的网络结构对图片不同尺度的特征信息进行提取.并结合内存优化,对网络的特征提取部分采用深度可分离卷积,使得网络相对于ResNet参数总量下降1/3.同时文中对特征解码器进行设计,网络计算中跳跃连接传递的参数量对比ResNet下降了68.61%.本文的轻量化深度估计网络参数位宽由32比特降至3比特.实验结果表明,轻量化后的深度估计网络的网络参数大小下降90.59%,在KITTI数据集上绝对相对误差为16.0%,最终轻量化的网络大小从34.12MB下降到了3.21MB.

A large number of proposed monocular depth estimation network research has a huge and bloated network structure, which will have the problems of large occupation and high delay in the actual deployment.In order to solve the problems, a lightweight depth estimation network based on learned step quantization strategy is proposed in this paper.The network adopts the structure of feature pyramid network(FPN)to extract the feature information of different scales.Combined with memory optimization, depthwise separable convolution is used in the feature extraction part of the network, so that the total amount of network parameters relative to ResNet is reduced by 1/3.At the same time, the skip connection transfer parameters in network computing is reduced by 68.61% compared with ResNet due the fine designed decoder.In this paper, the lightweight depth estimation network parameter bit width is reduced from 32 bits to 3 bits.The experimental results show that the network parameter size of the lightweight depth estimation network is reduced by 90.59%,and the absolute relative error on the KITTI data set is 16.0%.Finally, the lightweight network size is reduced from 34.12 MB to 3.21 MB.