DIRBW-Net: An Improved Inverted Residual Network Model for Underwater Image Enhancement

Abstract

Underwater photography is challenged by optical distortions caused by water absorption and scattering phenomena. These distortions manifest as color aberrations, image blurring, and reduced contrast in underwater scenes. To address these issues, this paper proposes a novel underwater image enhancement model, called DIRBW-Net, leveraging an improved inverted residual network. In order to minimize the interference of the Batch Normalization (BN) layer on color information, newly designed Double-layer Inverted Residual Blocks (DIRBs) are introduced, which omit the BN layer and extract deep feature information from the input images. Subsequently, each input image is fused with the intermediate feature map using skip connections to ensure consistency between local and global image information, thus effectively enhancing the image quality. In the concluding phase, effects of diverse activation functions are studied, opting for the h-swish activation function to further boost the overall model performance. DIRBW-Net is evaluated on a public dataset, with comparisons drawn against existing representative models. The experiments showcase a notable success in enhancing the underwater image quality when using the proposed model.

Keywords

• Underwater image enhancement
• convolutional neural network (CNN)
• residual network
• deep learning