Image dehazing algorithm based on multi-scale feature optimization in encoding and decoding
The uneven distribution of real-world haze often leads to poor recovery quality of imagescaptured under actual haze conditions by networks trained on synthetic datasets. Additionally,the large number of parameters in existing dehazing models affects the real-time performance of haze removal. To address these two issues,this paper proposes an image dehazing algorithm based on multi-scale feature optimization in encoding and decoding with a reduced number of parameters to ef-fectively remove haze from real-world images. Firstly,the encoding part employs cross-channel con-textual attention to implicitly model the relationships between pixels,thereby preserving the structure of objects in the dehazed images. Then,an information tuning subnet is designed to compensate for the shallow information missed by the encoder,addressing the problem of coarse detail recovery. Finally,the decoding part incorporates a feature correction subnet that utilizes a subtractive residual structure toreduce noise and ensure the accuracy of the output. Experiments on various real-world haze datasets demonstrate the generalizability of the proposed method. The results show that on the REVIDE real-world haze dataset,the proposed method achieves a 1.25 dB improvement in PSNR compared to the MSBDN method,with a 46％ reduction in the number of parameters. Furthermore,on the O-Haze,I-Haze,and RTTS real-world indoor and outdoor haze test sets,the proposed method outperforms other dehazing methods in terms of PSNR and visual quality.
signal and information processingimage dehazingdeep learningreal hazeencoding and decoding
万方数据
维普
