Dehazing Model Based on Bidirectional Encoder Representations and Heterogeneous Data

After two years of data collection, we built a dataset named Taiwan Weather Datasets with Surveillance Image (TWDSI), including image data and numeric data from weather and air quality monitoring stations. Dehazing, defogging, and de-raining algorithms were image processing techniques to remove weather or atmospheric effects from lowvisibility images. Conventional, these frameworks consider the crucial information to recover images exhibited in the images themself. However, taking dehazing algorithms as an example, the haze level was highly correlated to air quality; thus, data obtained from corresponding monitoring stations mattered. In the past, these kinds of research were limited because of the difficulty of simultaneously collecting real-world images, weather, and air quality data. To our best knowledge, this is the first study attempting to gather images and numeric data to fuse and recover visibility from a real-world image. In our framework, we did not label images with weather labels; instead, we designed a BERT-like model to fuse the data in latent spaces so that the data was naturally clustered in the latent space. We also tested our method and compared its performance with some state-of-the-art models. As a result, our model outperforms others, showing that using our framework has a high potential for development.