Image transformation using super resolution

Abstract

This study aimed to evaluate the effectiveness of two super resolution imaging (SR) models based on residual, Enhanced Deep Residual Networks for Single Image Super-Resolution (EDSR) and Densely Residual Laplacian Super-Resolution (DRLN), in enhancing the quality of photographic images. Following standard practice for model validation on the SR problem, quality was measured by computing objective metrics relative to benchmark image datasets containing full-resolution reference images. Both a standard benchmark of diverse, high-quality images (DIV2K) and a domain-specific image corpus (BeeMachine) were used for this purpose. One specific aim of SR is to improve outdoor images of organisms in the wild, for computer vision tasks such as object detection. This was the motivation for using BeeMachine image corpus, consisting of crowdsourced bumble bee photographs. Photographs from this collection vary in degree of motion blurring, viewing distance, and focal distance, to a greater degree than in standard SR benchmarks. To address this challenge, two general-purpose deep learning models were applied: Enhanced Deep Super Resolution (EDSR) and Densely Residual Laplacian Networks (DRLN). Experiments were conducted to compare EDSR and DRLM performance using (DIV2K) as a validation standard and BeeMachine to assess the effectiveness of SR on a specific ecologically-valid use case. Results of these models were examined using a set of objective metrics such as peak signal-to-noise ratio (PSNR) and structural similarity (SSIM). These results demonstrated that both models were effective at increasing image resolution. When compared to the EDSR model, the DRLN model achieved a higher PSNR and SSIM value. Furthermore, the models were also tested on images with varying levels of degradation to evaluate their robustness. The results showed that while both models performed well on moderately degraded images, the DRLN model was better at handling highly degraded images. This suggests that the DRLN model may be more suitable for enhancing the quality of bumble bee images in challenging field conditions where image quality is compromised. This study thus showed that extant models (EDSR and DRLN) whose performance is competitive with the state-of-the-art on an idealized general-purpose corpus (DIV2K) exhibit similar performance on the BeeMachine) corpus. This confirms the potential of SR using these models to enhance uploaded photographs at the lower-resolution (half and quarter scale) end of the experimental range, and to thereby facilitate research on the behavior and ecology of these important pollinators. While both models performed well, the DRLN model was found to be more effective in handling highly degraded images.