Single image dehazing has become an important vision task for prevailing image degradation caused by detrimental atmosphere transmission conditions. Attention mechanism has been widely utilized in learning based methods to assist the model to discard redundant information and hence boost the performance. However, existing methods are mainly dealing with either channel-wise or pixel-wise attention, which requires much more parameters when the size of feature maps increases. In this paper, we introduce the operation-wise attention in the proposed Feedback Operation-Attention Network (FOANet) to focus on attaining optimal combination of network operations for image dehazing. Specifically, our model consists of two main steps. First, the extracted features of hazy input image are fed to the novel operation-attention block which can adjust the weight of different operations dynamically to produce the optimal processed features. The operation space of an operation-attention block comprises vanilla and dilated convolutions with different kernel sizes along with max pooling and average pooling. Second, we adopt curriculum learning with feedback mechanism to continually refine the features in a recurrent fashion and generate the haze-free image. The experimental results on both synthetic and realistic subsets of RESIDE dataset have demonstrated our method can perform dehazing favorably against other dehazing algorithms.
|APSIPA Transactions on Signal and Information Processing
|Published - 2023