隐式神经表征(Implicit Neural Representation, INR)是近年来广受关注的一种数据表示形式。与传统的基于网格的离散表示方法不同，该隐式表征方法利用强大的神经网络模型，学习数据内部的非线性映射，从而实现对数据的隐式编码。由于其紧凑的参数化存储模式、强大的非线性表征能力以及天然适用于逆问题求解等优点，隐式神经表征在多媒体数据压缩、物理驱动的场模拟、三维场景的表示等领域发挥着重要的作用。然而，该表征形式的拟合编码过程会严重消耗算力资源和时间，成为了其进一步应用的主要障碍。为了解决这一挑战，本文提出了两种通用的隐式神经编码的加速方法：(1) 基于对称幂变换的加速方法。该方法提出了"范围限定性-对称性''假设，并基于此假设设计了一种可逆的非线性对称幂变换来加速隐式神经表征的编码过程。该变换同时满足范围限定和对称性特性，并通过偏差感知校正和自适应软边界等技术解决了极端偏差放大和连续性破坏性的问题。实验证明该方法可显著提升隐式表征在各类数据上的训练效率。(2) 基于进化算法启发的坐标采样加速方法。该方法受进化算法启发，将训练样本视为种群个体，设计了稀疏适应度评估函数、频率引导的杂交算法和无偏变异等机制。与传统隐式神经表征训练全量坐标点的方法相比，该方法可以在48%-66%的训练时间缩减下，仍能确保优秀的收敛性能。该方法在同期基于采样的加速策略中达到了领先水平，显著提升隐式神经表征的训练效率。以上两种方法分别从数据变换和采样策略的角度，提出了通用的隐式神经编码加速技术，可广泛应用于音频、图像、视频以及三维表示等拟合任务，可适配于各种隐式神经架构和已有加速方法。该研究为进一步推广隐式神经表征在各类应用中的使用，提供了有益的探索和解决思路。
Implicit Neural Representation (INR) has emerged as a novel data representation in recent years. Unlike traditional grid-based discrete representations, this implicit representation method leverages powerful neural network models to learn the complex nonlinear mapping within data, achieving implicit encoding. Due to its compact parametric storage and natural friendliness to inverse problems, INR plays an important role in various domains such as multimedia data compression, physics-driven field simulation, and 3D scene representation. However, the encoding process of this representation is severely constrained by its heavy consumption of computational time and resources, which has become the main obstacle to its further deployment. To address this challenge, this thesis proposes two universal acceleration methods for implicit neural representations:(1) Transformation-based Method: Symmetric Power Transformation. This method proposes the “Range-Defined Symmetry” hypothesis and designs a reversible nonlinear symmetric power transformation to accelerate the implicit neural encoding process. The method incorporates deviation-aware calibration and adaptive soft boundary techniques to enhance robustness. Experiments demonstrate that this method can significantly improve the training e?iciency of implicit representations for various data form.(2) Sampling-based Method: Evolutionary Selector. Inspired by Evolutionary Algorithms, this method treats training samples as population individuals and designs sparse fitness evaluation, frequency-guided crossover, and unbiased mutation mechanisms. Compared to conventional full-sample INR training, this method can achieve 48%-66% training time reduction while ensuring superior convergence, establishing the state-of-the-art in sampling-based acceleration.The above two methods respectively tackle the challenges from data transformation and sampling strategy, proposing universal acceleration techniques for implicit neural encoding that can be broadly applied to audio, image, video, and 3D representation fitting tasks. This research provides valuable exploration and solutions to further promote the deployment of INR in various applications.
