Adaptive feature fusion network based on boosted attention mechanism for single image dehazing
Zhao Wang,Feng Li,Runmin Cong,Huihui Bai,Yao Zhao
DOI: https://doi.org/10.1007/s11042-022-12151-4
IF: 2.577
2022-02-17
Multimedia Tools and Applications
Abstract:Recently convolutional neural networks based methods have achieved significant improvements in image dehazing. However, these algorithms still face the challenge of producing haze-free images while preserving credible contrast and color fidelity. In this paper, we propose an adaptive feature fusion network to remove the haze and ensure realistic details from global and local perspectives. On the global scale, we learn compact feature representations by progressive downsampling, which can provide the overall information from the encoded high-level semantic context. Besides, dilated convolution is adopted to expand the receptive fields, which can effectively capture the contextual information and alleviate the details loss of resolution reduction. Correspondingly, the proposed method employs a local branch to enrich the feature representations and further emphasize the meaningful information for image details recovery. To this end, we design the residual dense attention block (RDAB) which encourages mid-level feature aggregation and persist memory by dense connections. Within the RDAB, a boosted attention mechanism (BAM) is presented to explicitly model the feature interdependencies between channels under different scales. Then, a weighting operation is conducted to balance the information flow received from these scales. Moreover, an adaptive weighted network is developed to achieve a good trade-off between the contributions of global and local information for semantical image dehazing. To take full account of quality evaluation, we use the L1 smooth loss and perceptual loss to reconsturct the dehazed images. Extensive evaluation demonstrates the superior performance of our method is 4db PSNR values and 0.1 SSIM value more than the related work while preserving credible contrast and color fidelity.
computer science, information systems, theory & methods,engineering, electrical & electronic, software engineering