Multi‐scale GAN with residual image learning for removing heterogeneous blur

Abstract

Processing images with heterogeneous blur remains challenging due to multiple degradation aspects that could affect structural properties. This study proposes a deep learning‐based multi‐scaled generative adversarial network (GAN) with residual image learning to process variant and in‐variant blur. Different scaled images with corresponding gradients are concatenated as a multi‐channel single input for the proposed GAN. Residual‐ and dense‐networks are combined to explore salient features in the bottleneck section while addressing the vanishing gradient problem. A hybrid content loss function with a gradient penalty minimises the error between generated and ground truth images. Due to structure sparsity, the generated output may lose some information that leads to artifacts. Residual image learning with dilation and end‐to‐end training is used to resolve this issue by recovering high‐resolution anatomical details. Three different datasets: GoPro, Köhler, and Lai, with variant and in‐variant blur, are used to perform qualitative and quantitative analyses. Experiments show the proposed method is effective in reducing blur while preserving structural properties compared to multiple preprocessing techniques for image analysis. Moreover, the consistently improved performance over multiple publicly available datasets validates the merits of the proposed method for large data analysis.

Improving the quality of light‐field data extracted from a hologram using deep learning

“…A GAN-based deep-learning model [13][14][15][16][17] was used to reduce speckle noise and improve the spatial resolution of orthographic view images extracted from holograms. The GAN model is a generative model that produces realistic images and human-like text.…”

“…In the proposed method, holograms are generated using two-dimensional image data for deep-learning model training, and a dataset is constructed by applying bandpass filtering. A generative adversarial network (GAN)-based deep-learning model [13][14][15][16][17] was designed to improve the quality of the extracted light-field data and was trained using the dataset. Finally, using the trained model, light-field data with reduced speckle noise and improved spatial resolution were obtained.…”

Unsupervised Multi-Scale Image Enhancement using Generative Deep Learning Approach

“…It has been observed that a captured image of a real scene or object can be degraded by inaccurate capture and optical factors such as improper depth of field, poor focus, camera shake, object movement, short exposure, and poor optical quality. Tasks involving image identification or classification can be impeded by noisy visuals with imperfections [2]. Noise often comes out in images as detached pixels or blocks of pixels, causing heavy visual effects.…”