A Modified GCANet Image Dehazing Algorithm for Road Scenes

Environmental perception algorithm is one of the key technologies for driverless car, haze weather may greatly reduce the accuracy of object detection, bringing danger for this case. In this paper, we propose a modified image dehazing algorithm based on an excellent dehazing algorithm, i.e., GCANet. The modified dehazing algorithm presents the multiscale feature enhancement strategy, which adds the residual structure into the decoder module to fuse and enhance multiscale features. It also presents the pixel-level haze component association strategy, which extracts the information of adjacent pixels to help the prediction on haze component. Experiments show that both the two strategies can enhance the dehazing performance by itself, and the fusion of them can obtain a better dehazing performance. And compared with the mainstream dehazing algorithms, the modified GCANet achieves the optimal dehazing effect under the same experimental condition. It can output a high-quality dehazing image which is beneficial to the subsequent tasks of driverless vision perception algorithm.