【摘要】 自然水体成像中湍流及悬浮颗粒等环境因素会造成水下采集的图像存在扭曲失真、分辨率低、背景模糊等问题,为了解决上述问题并进一步提高图像重建和复原的质量,提出了一种改进的基于残差网络的图像超分辨率重建方法,该方法将网络中的残差密集块和自适应机制进行融合,有效解决深度学习网络中经常遇到的梯度爆炸问题,同时能够抑制无用信息的学习,充分利用重要特征信息。为了使网络适应水下噪声环境,通过自建水下系统对目标板分别在清水中和浑浊微湍流水域中进行采集并对其进行图像配对生成训练对,并在河流和海洋水域下采集图像生成测试集。实验结果表明,在微湍流的海洋水域和河流水域中,较传统的水下图像处理和神经网络算法,使用改进的残差网络算法能够很好地对水下图像进行重建,重建图像的边缘信息得到了极大的保留,图像的重建效果更好。更多还原

【Abstract】 Natural environment factors such as turbulence and suspended particles in water imaging can cause image distortion,low resolution,and fuzzy background of underwater acquisition.In order to solve the above problems and further improve the quality of image reconstruction and rehabilitation,this paper puts forward an improved image super-resolution reconstruction based on residual network method.This method will in residual dense network of fusion and adaptive mechanism,effectively solve the deep learning gradient explosion problems often encountered in network,also can inhibit learning of useless information,make full use of the important feature information.In order to adapt the network to the underwater noise environment,a self-built underwater system is used to collect the target plate in clear water and turbid micro-turbulent waters respectively,and the training pair of image generation is performed on the target plate,and the test set of image generation is collected under rivers and ocean waters.The experimental results show that in the micro-turbulent ocean and river waters,compared with the traditional underwater image processing and neural network algorithm,the improved residual network algorithm can reconstruct the underwater ima-ge very well.更多还原