Multiscale fusion dehazing algorithm based on channel attention mechanism with sub-pixel convolution

This study aimed to propose a multiscale fusion dehazing algorithm based on the channel attention mechanism with sub-pixel convolution (MCASDA) to address the issues of blurred image details, colour distortion, and artefact residue in complex haze scenes. The algorithm was based on the convolutional networks for biomedical image segmentation architecture and achieved multilevel feature co-optimisation through encoder and decoder frameworks. The enhanced residual network was designed with the convolutional layer of the Swish activation function in the encoder stage. Combined with the channel attention mechanism, thus significantly improving the ability of the network to focus on key features. Sub-pixel convolutional upsampling was used to replace the traditional transpose convolution in the decoder stage, avoiding checkerboard artefacts through pixel rearrangement. Moreover, the multiscale feature fusion module was combined with back-projection feedback to recover high-frequency details and spatial information effectively. The evaluation outcomes indicated that MCASDA achieved notable improvements in peak signal-to-noise ratio and structural similarity index on the synthetic objective testing set dataset (29.09 dB and 0.9632, respectively), surpassing conventional approaches. These findings substantiated the algorithm's superior performance in dehazing, thereby effectively producing higher-quality dehazing images with enhanced clarity.