Hu Xianjie, Liu Jing, Li Heng, Liu Hui, Xue Xiaojun
Faculty of Information Engineering and Automation, Kunming University of Science and Technology, Kunming, China.
PeerJ Comput Sci. 2024 Apr 30;10:e1783. doi: 10.7717/peerj-cs.1783. eCollection 2024.
Underwater images suffer from color shift, low contrast, and blurred details as a result of the absorption and scattering of light in the water. Degraded quality images can significantly interfere with underwater vision tasks. The existing data-driven based underwater image enhancement methods fail to sufficiently consider the impact related to the inconsistent attenuation of spatial areas and the degradation of color channel information. In addition, the dataset used for model training is small in scale and monotonous in the scene. Therefore, our approach solves the problem from two aspects of the network architecture design and the training dataset. We proposed a fusion attention block that integrate the non-local modeling ability of the Swin Transformer block into the local modeling ability of the residual convolution layer. Importantly, it can adaptively fuse non-local and local features carrying channel attention. Moreover, we synthesize underwater images with multiple water body types and different degradations using the underwater imaging model and adjusting the degradation parameters. There are also perceptual loss functions introduced to improve image vision. Experiments on synthetic and real-world underwater images have shown that our method is superior. Thus, our network is suitable for practical applications.
由于光在水中的吸收和散射，水下图像会出现颜色偏移、对比度低和细节模糊等问题。质量退化的图像会严重干扰水下视觉任务。现有的基于数据驱动的水下图像增强方法未能充分考虑与空间区域衰减不一致以及颜色通道信息退化相关的影响。此外，用于模型训练的数据集规模较小且场景单一。因此，我们的方法从网络架构设计和训练数据集两个方面解决了这个问题。我们提出了一种融合注意力块，将Swin Transformer块的非局部建模能力集成到残差卷积层的局部建模能力中。重要的是，它可以自适应地融合携带通道注意力的非局部和局部特征。此外，我们使用水下成像模型并调整退化参数来合成具有多种水体类型和不同退化程度的水下图像。还引入了感知损失函数来改善图像视觉效果。在合成和真实水下图像上的实验表明，我们的方法具有优越性。因此，我们的网络适用于实际应用。
