«上一篇/Previous Article|本期目录/Table of Contents|下一篇/Next Article»

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:

期数:

2025年10期

页码:

43-52

栏目:

媒体计算

出版日期:

2025-10-10

文章信息/Info

Title:

Deep Robust Image Reconstruction Based on Autoencoder and Sparse Learning

文章编号:

1673-629X(2025)10-0043-10

作者:

张俊洋; 明镝

重庆理工大学计算机科学与工程学院,重庆 400054

Author(s):

ZHANG Jun-yang; MING Di

School of Computer Science & Engineering,Chongqing University of Technology,Chongqing 400054,China

关键词:

图像重构; 深度鲁棒模型; 无监督学习; 平滑L1自编码器; 稀疏特征学习; 近端梯度下降

Keywords:

image reconstruction; deep robust models; unsupervised learning; smoothed L1 autoencoder; sparse feature learning; proximal gradient descent

分类号:

TP391.41

DOI:

10.20165/j.cnki.ISSN1673-629X.2025.0138

摘要:

为了解决现有自编码器在噪声环境下由于L2损失和L2正则化导致图像重构效果不佳的问题,该文提出了一种新颖的基于平滑L1自编码器和稀疏特征学习的深度鲁棒图像重构方法(SL1AE-SFL),能够显著提升图像重构的性能和效果。首先,该文提出了平滑L1自编码器,采用L1损失函数和平滑sReLU激活函数,替代传统的L2损失函数和ReLU激活函数,从而提升了模型对噪声和异常数据的鲁棒性。其次,针对L2正则化难以有效抑制噪声特征影响的问题,该文提出了基于稀疏特征学习的鲁棒自编码器方法,包括SL1AE-SFL(L1)和SL1AE-SFL(L21),通过施加基于L1和L21范式的稀疏正则化约束,使得模型不被噪声特征所干扰、更加关注重要特征,从而显著增强了重构的准确性。最后,该文提出了一种基于近端梯度下降的高效优化算法,有效缓解了L1和L21稀疏正则化在训练过程中的低效问题。在多个数据集上的实验结果表明,SL1AE-SFL相较于其他方法在重构误差和聚类精度方面均表现出显著优势。此外,还分析了网络结构对重构质量的影响,消融实验表明,适当增加网络层数可以进一步提升图像重构效果,验证了深层网络在特征提取与图像重构方面的有效性。

Abstract:

To address the issue of poor image reconstruction performance in noisy environments caused by L2 loss and L2 regularization in existing autoencoders,we propose a novel deep robust image reconstruction method based on a Smoothed L1 Autoencoder and Sparse Feature Learning (SL1AE-SFL), which significantly enhances reconstruction quality and performance. Firstly, we introduce the Smoothed L1 Autoencoder,which replaces the traditional L2 loss function and ReLU activation function with the L1 loss function and smoothed sReLU activation function,thereby improving the model’s robustness against noises and outliers. Secondly,to tackle the challenge of L2 regularization failing to effectively suppress the influence of noisy features,we propose a robust autoencoder approach based on sparse feature learning,including SL1AE-SFL(L1) and SL1AE-SFL(L21). By imposing sparse regularization constraints based on L1 and L21 norms,the model becomes less affected by noisy features and focuses more on essential features, thereby significantly improving reconstruction accuracy. Finally,we propose an efficient optimization algorithm based on proximal gradient descent,which effectively alleviates the training inefficiency of L1-norm and L21-norm based sparse regularizations. Experimental results on multiple datasets demonstrate that SL1AE-SFL outperforms other methods in terms of reconstruction error and clustering accuracy. Moreover,we analyze the impact of network architecture on reconstruction quality,and ablation studies show that appropriately increasing the number of network layers can further enhance image reconstruction performance,validating the effectiveness of deep networks in feature extraction and image reconstruction.

常用功能

更新日期/Last Update: 2025-10-10