The Paradigm Shift in Hyperspectral Image Compression: A Neural Video Representation Methodology

作者：Nan Zhao, Tianpeng Pan, Zhitong Li, Enke Chen, Lili Zhang · 发表于：Remote Sensing · 年份：2025 · DOI：10.3390/rs17040679 · 被引用次数：6 · 研究领域：Advanced Data Compression Techniques、Image and Signal Denoising Methods、Image Retrieval and Classification Techniques

In recent years, with the continuous development of deep learning, the scope of neural networks that can be expressed is becoming wider and their expressive ability stronger. Traditional deep learning methods based on extracting latent representations have achieved satisfactory results. However, in the field of hyperspectral image compression, the high computational cost and the degradation of their generalization ability reduce their application. We analyze the objective formulation of traditional learning-based methods and draw the conclusion that rather than treating the hyperspectral image as an entire tensor to extract the latent representation, it is preferred to view it as a stream of video data, where each spectral band represents a frame of information and variances between spectral bands represent transformations between frames. Moreover, in order to compress the hyperspectral image of this video representation, neural video representation that decouples the spectral and spatial dimensions from each other for representation learning is employed so that the information about the data is preserved in the neural network parameters. Specifically, the network utilizes the spectral band index and the spatial coordinate index encoded with positional encoding as its input to perform network overfitting, which can output the image information of the corresponding spectral band based on the index of that spectral band. The experimental results indicate that the proposed metho...