基于MobileNet的轻量级网络构建及其超分辨率图像重建

Constructing lightweight neural networks based on MobileNet to reconstruct single image super-resolution

目的:研究轻量级网络的超分辨率重建。方法:尝试在图像超分辨率重建中引入MobileNet网络结构,并使用MobileNet v2网络结构对网络进行改进。结果:通过将标准的卷积网络分解为深度卷积和逐点卷积操作,该网络将参数数量和计算量缩减为原来的1/4左右。结果显示除了在扩大因子为×2的情况下重建效果有所下降之外,在其他的尺度上都取得了更好的效果。使用MobileNet v2网络结构对网络改进以后,该网络能在参数数量和计算量增加不多的情况下进一步提升效果,重建效果超过所有对比方法。结论:所构建的两个轻量级网络不仅在定性指标上面有更好的结果,而且在视觉效果上也具一定优势。

Aims: This paper aims to study the image super-resolution of lightweight neural networks. Methods:The network structure of MobileNet was introduced into image super-resolution reconstruction to improve the network by using the network structure of MobileNet V2. Results: By decomposing the standard convolution network into depthwise convolution and pointwise convolution operation, the number of parameters and computation was reduced to about 1/4 of the original one. The results showed that the reconstruction effect decreased when the scale factor was ×2 and better results were achieved on other scales. Improved by using MobileNet V2, the network increased the effect with little increase of parameters and computation amount, having the best reconstruction effect. Conclusions: This two lightweight networks not only have better results in qualitative indexes, but also they have some advantages in visual effects.