HSI-IPGAN: Hyperspectral Image Inpainting via Generative Adversarial Network

Abstract

Due to the instability of the hyperspectral imaging system and the atmospheric interference, hyperspectral images (HSIs) often suffer from losing the image information of areas with different shapes, which significantly degrades the data quality and further limits the effectiveness of methods for subsequent tasks. Although mainstream deep learning-based methods have achieved promising inpainting performance, the complicated ground object distributions increase the difficulty of HSIs inpainting in practice. In addition, spectral redundancy and complex texture details are two main challenges for deep neural network-based inpainting methods. To address the above issues, we propose a novel inpainting model based on generative adversarial networks for HSI (HSI-IPGAN). To reduce the redundancy of spectral information, a multi-frequency channel attention module is designed to strengthen the essential channels and suppress the less significant ones, which calculates adaptive weight coefficients by converting feature maps to the frequency domain. Furthermore, we propose to constrain the generation of missing areas from both global and local perspectives, by fully leveraging the HSI texture information, so that the overall structure information and regional texture consistency of the original HSI can be maintained. The proposed method has been extensivelyevaluated on the Indian Pines and FCH datasets. The promising results indicate that HSI-IPGAN outperforms the state-of-the-art methods in quantitative and qualitative assessments.

ENTS: An Edge-native Task Scheduling System for Collaborative Edge Computing

“…Other works [13], [14] tend to leverage the cloud-edge collaboration to partition and distribute the massive computation workload of LLM inference and finetuning. Wang et al [13] increase the throughput by distributing the computation between cloud servers and edge devices, and reducing the communication overhead of transmitting the activations between the central cloud and edge devices by leveraging the low-rank property of residual activations.…”

“…Wang et al [13] increase the throughput by distributing the computation between cloud servers and edge devices, and reducing the communication overhead of transmitting the activations between the central cloud and edge devices by leveraging the low-rank property of residual activations. Chen et al [14] efficiently leverage location-based information of edge devices for personalized prompt completion during collaborative edgecloud LLM serving. However, the latency between edge devices and the central cloud is usually high and unstable, which will affect the inference and finetuning performance of LLM.…”

“…However, they often lead to accuracy loss. Other works tend to use cloud-edge collaboration [13], [14], which partitions the LLMs into two sub-models and offloads part of the computation workload to the powerful cloud servers with high-end GPUs. However, the latency between edge devices and cloud servers is usually high and unstable.…”

A Unet-like Low-cost Image Inpainting Method Built upon Efficient ViT

“…The fundamental concept behind deep learning-based approaches is to acquire a mapping function that transforms a corrupted image into an unblemished one through a convolutional neural network (CNN) using an extensive corpus of training data. Further enhancement of images is accomplished via Generative Adversarial Networks (GANs) [30], the edgeconnect model [32], attention mechanism [5] and multi-scale feature fusion module [27].However, there remain various challenges, including managing intricate scenes, maintaining semantic and structural coherence, creating diverse and adjustable outcomes, and objectively and subjectively assessing performance. This overview of the advancement of image inpainting seeks to be a valuable resource, stimulating future research and innovation in this area.…”