研究目的
To restore the visual quality of underwater images without requiring ground truth data by using a set of image quality metrics to guide the restoration learning process.
研究成果
The proposed methodology improves the visual quality of underwater images by not degrading their edges and performs well on the UCIQE metric. It offers a physically-plausible restoration of the input with better quality than other approaches.
研究不足
The method relies on the quality of the image quality metrics used to guide the restoration process. The effectiveness of the restoration is limited by the accuracy of these metrics in representing human visual perception.
1：Experimental Design and Method Selection:
The methodology comprises two-phase learning. Initially, supervised training is performed by fine-tuning the DehazeNet to estimate the transmission map. Then, the input image is restored according to an image formation model. In the second phase, a loss function composed of quality metrics is minimized to perform image restoration.
2：Sample Selection and Data Sources:
For the supervised phase, a set of synthetic underwater images is created using the physically based rendering engine PBRT. In the unsupervised phase, three datasets are used: images from the underwater-related scene categories of the SUN dataset, images from a water tank with different turbidity levels, and a subset of images from the work of Ancuti et al.
3：List of Experimental Equipment and Materials:
A convolutional network with three convolutional layers is used. The images are normalized for values between [0,1] after the restoration.
4：Experimental Procedures and Operational Workflow:
The network is fine-tuned using synthetic images, then switched to unsupervised mode using real-scene images. The weights of the network are updated by propagating the IQM error backward.
5：Data Analysis Methods:
The UCIQE metric and border integrity are used for evaluation.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
