A multi‐head convolutional attention residual network for prepress image super‐resolution reconstruction in digital printing

Abstract For images intended for digital printing production, besides vector graphics that are resolution‐independent yet typically simplistic in colour and structure, there is a consistent and pressing need for bitmap images of natural scenes that possess abundant detail. However, most of the existing bitmap images do not meet the precision standards required for digital printing, leading to a compromise in the visual quality of the products. Deep learning‐based super‐resolution reconstruction techniques can aid in restoring more refined textural features of images, thereby enhancing the quality of digital printing outcomes. In this paper, a multi‐head convolutional attention (MHCA)‐based residual network is proposed for prepress image super‐resolution reconstruction in digital printing. First, we remove the batch normalisation layers from the deep feature extraction module, which enhances the network's flexibility and reduces graphics processing unit memory usage. Then a new residual block named MHCARes is designed by introducing the MHCA mechanism, to focus on high‐frequency information from different channels and spatial dimensions. Furthermore, a flexible and efficient reconstruction module that can make the network suitable for non‐exponentially and slowly increasing scale factors is introduced. And the network is finally trained by the Charbonnier loss function. Experimental comparisons are conducted between the proposed method and other state‐of‐the‐art methods. The results show that the proposed method exhibits superior performance in terms of peak signal‐to‐noise ratio and structural similarity index, and the reconstructed digital printing images possess more realistic fine textures, meeting the needs of digital printing production.