Proceedings Article | 20 January 2023
KEYWORDS: Image segmentation, Super resolution, Network architectures, Data modeling, Image fusion, Feature extraction, Image restoration, Image enhancement, Neural networks, Image processing
To address the problem of low accuracy of underwater image segmentation images due to low resolution, low visibility, a wide variety of objects and insufficient illumination of underwater acquired images, we build a super-resolution underwater image segmentation network based on multi-scale fusion. We apply Resnet50 as the backbone feature extraction network to the U-shaped network architecture, which can well enhance the feature capture capability of the model, and apply the upsampling output of the feature extraction model combined with the super-resolution module to obtain a refined output. Compared with other networks, our model has significant segmentation capability on the SUIM underwater image segmentation dataset. The shallow feature information of the input image can be better restored and preserved, and the edge feature information of the image is better preserved with higher accuracy, and the model has higher cross-merge ratio and accuracy. Our work is applied to the SUIM underwater image segmentation dataset and compared with U-Net( vgg16),U-Net(Resnet50), PSPNet(Mobilenet), PSPNet(Resnet50), deeplab(mobilenet) networks on this dataset, and our proposed network The average cross-merge ratio (MIou) of our proposed networks improved by 4.29%, 1.58%, 7.44%, 2.69% and 3.24%, respectively, and the pixel accuracy (mPa) improved by 3.24%, 1.57%, 5.4%, 1.93% and 5.73%, respectively.
