基于蓝图可分离卷积的轻量级水下图像超分辨率重建

Super-Resolution Reconstruction of Lightweight Underwater Images Based on Blueprint Separable Convolution

摘要:

由于水体中存在的悬浮颗粒以及高频随机运动的湍流引起光的散射和吸收而导致水下图像存在纹理模糊、分辨率低、扭曲失真等系列问题，而目前存在的大部分深度学习图像超分辨率重建算法存在着计算复杂、模型的复杂度大、内存占用高等不足。针对这些不足，提出基于蓝图可分离卷积的轻量级水下图像超分辨率重建网络，该模型分为浅层特征提取、深度特征提取、多层特征融合以及图像重建四个阶段，深度特征提取阶段中，在BSRN的基础上去除特征蒸馏分支、采用增加通道数进行补偿，同时利用三个蓝图卷积来进行残差局部特征学习以简化特征聚合，实现网络的轻量化。实验结果表明，所提出的方法在运行时间、参数量、模型复杂度方面均优于目前已提出的超分算法，放大因子为2和4时，峰值信噪比(PSNR)和结构相似度(SSIM)均值分别达到了31.5560dB、0.8620和27.7088dB、0.7213，重建质量获得进一步提升。

Abstract:

Abstract: Due to the scattering and absorption of light caused by the suspended particles in the water body and the turbulence of high-frequency random motion, the underwater image has a series of problems such as blurred texture, low resolution and distortion. However, most of the existing deep learning image super-resolution reconstruction algorithms have the problems of complex computation, large model complexity and high memory occupation. To solve this problem, a lightweight underwater image super-resolution reconstruction network based on blueprint separable convolution is proposed. The model is divided into four stages: shallow feature extraction, deep feature extraction, multi-layer feature fusion and image reconstruction, In the depth feature extraction stage, feature distillation branches are removed on the basis of BSRN and the number of channels is increased for compensation. At the same time, the three blueprints convolution is used to carry out residual local feature learning to simplify feature aggregation and realize the lightweight of the network. The experimental results show that the proposed method is superior to the currently proposed hyperspectral algorithm in terms of running time, parameter quantity and model complexity. When the amplification factor is 2 and 4, the peak signal-to-noise ratio (PSNR) and structure similarity (SSIM) average values reach 31.5560dB, 0.8620 and 27.7088dB, 0.7213 respectively, and the reconstruction quality is further improved.