%0 Journal Article
%T 多尺度注意力机制的水下图像增强算法<br>Underwater Image Enhancement Based on Multi-Scale Attention Mechanism
%A 王士川
%A 杜玲
%J Computer Science and Application
%P 2506-2516
%@ 2161-881X
%D 2021
%I Hans Publishing
%R 10.12677/CSA.2021.1110254
%X <div style=\"text-align:justify;\">
	<span style=\"font-size:14px;\">由于不同波长的光在水环境中的衰减差异，原始的水下图像会出现偏绿、偏蓝和白雾的问题。针对水下图像增强出现的上述问题，本文通过结合多尺度密集块和注意力机制，提出了一个生成式对抗网络。在生成器中使用多尺度残差密集块，在后端使用注意力机制块来关注前端的特征，前者用来提高网络学习能力，后者增强网络的特征选择能力。判别器的每一层都使用谱归一化，放宽判别标准，以减少异常数据引起的梯度崩溃。同时，结合多种损失函数来增强生成对抗网络的泛化能力。提出的算法在公共数据集上进行了测试，并在实际应用中得到了验证。实验结果表明，该算法比现有算法有更好的性能。<br />
Due to the attenuation difference of different wavelengths of light in the water environment, the original underwater images show greenish, bluish and white haze problems. For underwater image enhancement, this paper proposes a generative adversarial network by combining multi-scale dense blocks and attention mechanism. The residual multiscale dense block is used in the generator, and the attention mechanism block is used in the back-end to focus on the front-end features, which can improve the learning ability and enhance the network’s ability to select features, respectively. Each layer of the discriminator uses spectral normalization and relaxes the discriminant criteria to reduce the gradient collapse caused by anomalous data. At the same time, multiple loss functions are combined to enhance the generalization ability of GAN networks. The proposed algorithm is tested on public datasets and validated in practical applications. Experimental results show better performance than existing algorithms.</span> 
</div>
%K 水下图像增强，生成对抗网络，深度学习<br>Underwater Image Enhancement
%K Generative Adversarial Network
%K Deep Learning
%U http://www.hanspub.org/journal/PaperInformation.aspx?PaperID=45733