FISRCN: a single small-sized image super-resolution convolutional neural network by using edge detection

作者：Kong, Luoyi; Wang, Fengbin^*; Yang, Fengyu; Leng, Lu; Zhang, Haotian

来源：MULTIMEDIA TOOLS AND APPLICATIONS, 2024, 83(7): 19609-19627.

DOI：10.1007/s11042-023-15380-3

• Single image super-resolution
• Small-sized image
• Pixel shuffle
• Sobel filter
• Median filter

In recent years, deep neural network-based models have shown remarkable success in achieving high-quality reconstruction for single image super-resolution. Among these models, generative adversarial network-based methods have produced satisfying results. However, these methods can be distorted when dealing with small-sized images containing a lot of interference information. This work proposes an improved convolutional neural network-based model (FISRCN) for super-resolution reconstruction of small-sized images, which can provide excellent colour restoration and rich texture features. In addition, to reconstruct the edge characteristics of small-sized images, a Sobel filter is introduced. Then two median filters are used to effectively smooth the noise. To improve runtime efficiency, pixel shuffle is used to upsample the image. The low-resolution image is then mapped to a high-resolution image using a convolutional neural network while simultaneously reconstructing the image using high-dimensional features. The experimental results demonstrate that FISRCN has a more balanced performance in terms of reconstruction quality and runtime efficiency (+ 1.295 dB in PSNR score, + 0.084 in SSIM score and -7.202 s in running time on small-sized images than Real-ESRGAN).

• 单位
  南昌航空大学