Skip to content

Exploring advanced autoencoder architectures for efficient data compression on EMNIST dataset, focusing on high-fidelity image reconstruction with minimal information loss. This project tests various encoder-decoder configurations to optimize performance metrics like MSE, SSIM, and PSNR, aiming to achieve near-lossless data compression.

License

Notifications You must be signed in to change notification settings

mdhabibi/Neural-Compression-with-Autoencoders

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

49 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

DataCompressNet

Project Overview: Data Compression with Complex Autoencoder Architectures

The primary objective of this project is to develop an efficient model for data compression. The focus is on leveraging complex autoencoder architectures to achieve significant dimensionality reduction while minimizing information loss. These models are adept at encoding high-dimensional data into a lower-dimensional space, effectively compressing the data while preserving its essential characteristics. This approach aims to reduce the memory footprint of data, which is crucial in various applications, including storage optimization and efficient data transmission. It's important to note that this form of compression primarily pertains to reducing dimensionality and memory usage, rather than decreasing storage size.

Autoencoder Architecture and Applications

*By Michela Massi - Own work, CC BY-SA 4.0, https://commons.wikimedia.org/w/index.php?curid=80177333*

Autoencoders are a type of neural network architecture used for unsupervised learning. They are designed to encode input data into a lower-dimensional representation and then reconstruct the original data from this representation. This process allows for learning efficient data codings in an automated fashion.

Applications of Autoencoders:

  • Data Compression: In the context of data compression, autoencoders excel at representing data in a more compact form, which helps in reducing memory usage without significantly compromising the data's integrity.
  • Dimensionality Reduction: Autoencoders can reduce the dimensionality of data, similar to Principal Component Analysis (PCA), making them useful for data visualization and noise reduction.
  • Feature Learning and Image Reconstruction Models: They are also used for automated feature learning, image reconstruction tasks like denoising and inpainting.

Design Choices for Encoder and Decoder:

For high-quality image reconstruction, I made specific design choices in the encoder and decoder parts of the autoencoder:

Encoder - I opted for Conv2D layers with strides instead of MaxPooling2D to:

  • Preservation of Spatial Information: By using only Conv2D layers with strides for downsampling, I aimed to preserve more spatial information compared to max pooling. Strided convolutions reduce dimensions while learning how to downsample, retaining more details that are crucial for accurate reconstruction.
  • Balancing Efficiency and Detail: I avoided MaxPooling2D to minimize information loss and ensure that the encoded representation retains sufficient detail for high-quality reconstruction.

Decoder - Conv2DTranspose over UpSampling:

  • Quality of Reconstruction: To reconstruct the images from the latent representation, I chose Conv2DTranspose layers over UpSampling2D. This decision was driven by the need for higher-quality image reconstruction.
  • Learnable Upsampling: Conv2DTranspose layers offer learnable parameters for upsampling, allowing the network to effectively learn how to expand the encoded representation back to the original image size.
  • Avoiding Artifacts: The use of Conv2DTranspose also helps in reducing artifacts that might occur with simpler upsampling methods.

Experimental Approach:

To determine the optimal layer configuration for the decoder, I tested four different models, each with a unique decoder structure. The performance of each model was assessed based on several key metrics. This evaluation was conducted for both the full autoencoder models and only their decoder parts, to gauge their respective abilities in reconstructing images accurately. This systematic exploration helps us understand the impact of decoder configurations on image reconstruction quality, providing insights for enhancing autoencoder performance in data compression tasks.

Clarification on Reconstruction Methods

Note that I employed two methods to reconstruct images from the EMNIST dataset as mentioned above: directly using the full autoencoder model and separately using the encoder and decoder models. While theoretically, these two approaches should yield identical results (as the encoder and decoder parts used separately are the same as those combined in the full autoencoder model), I included both methods in my experimentation for the following reasons:

  • Verification of Model Consistency: By reconstructing images using both methods, I could verify the consistency and correctness of my autoencoder architecture. This served as a practical check to ensure that the encoder and decoder parts function as intended when used both jointly (in the full autoencoder) and separately.
  • Illustrative Purposes: Demonstrating both methods provides a clearer understanding of the autoencoder's functionality, especially for those new to the concept. It illustrates the roles of the encoder and decoder, showing how data is first compressed and then reconstructed.
  • Foundation for Advanced Analysis: Separating the encoding and decoding processes lays the groundwork for more complex analyses, such as manipulating or analyzing the encoded representations (latent space) before decoding, which could be a subject for future work.
  • Educational Value: Including both methods adds educational value to the project, making it a more comprehensive resource for learners exploring different aspects of autoencoders.

Although including both methods might seem redundant, it was a deliberate choice to enhance the comprehensiveness of my analysis and to provide a robust demonstration of the autoencoder's capabilities in data compression and image reconstruction.

Dataset Used:

For this analysis, the EMNIST-Balanced dataset was chosen. It is a comprehensive dataset that includes a mix of digits, uppercase, and lowercase letters, offering a diverse range of characters for my models to learn and reconstruct. https://github.com/machinecurve/extra_keras_datasets

  • Below is a detailed analysis including explanations of each metric:

Metrics Explained:

  • Training Loss: Measures how well the model is performing during the training phase. Lower values indicate better performance.
  • Validation Loss: Indicates the model's performance on unseen data. It helps to identify overfitting.
  • Mean Squared Error (MSE): Quantifies the average squared difference between the reconstructed and original images. Lower MSE values signify more accurate reconstructions.
  • Structural Similarity Index (SSIM): Assesses the perceived quality of the reconstructed image compared to the original. Ranges from -1 to 1, with higher values indicating better image quality.
  • Peak Signal-to-Noise Ratio (PSNR): Measures the ratio between the maximum possible power of a signal and the power of corrupting noise. Expressed in decibels (dB), higher values represent better reconstruction quality.
  • Model01 demonstrates the best overall performance with the lowest training and validation loss, indicating excellent generalization. It also achieved the highest SSIM, reflecting superior image reconstruction quality.
  • Model02 shows a higher training and validation loss, suggesting potential overfitting. The lower SSIM and PSNR indicate a comparatively reduced ability in preserving image details during reconstruction.
  • Model03 offers a good balance between training and validation performance. It shows a high degree of image reconstruction accuracy, as evident from its SSIM and PSNR values.
  • Model04 falls between Models01 and 02 in terms of performance metrics. It displays a moderate level of image reconstruction capability, reflected in its SSIM and PSNR scores.

Summary

The evaluation metrics indicate that Model01 stands out as the most effective in reconstructing images from the EMNIST-Balanced dataset, with the highest SSIM and PSNR scores and the lowest loss values. Models03 and 04 offer competitive performance, with Model02 lagging slightly behind in terms of image quality preservation. These insights are crucial for further refinement of autoencoder models for image reconstruction tasks.

Data Compression Analysis

The original dataset, composed of uint8 grayscale images (28x28 pixels), occupied approximately 84.34 MB of memory. To prepare the data for the autoencoder, I converted it to float32 format, which increased the size to around 337.35 MB due to the higher precision and range of float32. The autoencoder model's encoder part successfully compressed this dataset, significantly reducing its size. The encoder output, a lower-dimensional representation of the data in float32 format, resulted in a compressed dataset of only 13.77 MB. This demonstrates the model's effectiveness in data compression, achieving a substantial reduction in size while potentially maintaining the integrity of the data for further processing or analysis.

  • Original Dataset Size (uint8): ~84.34 MB
  • Dataset Size after Conversion to float32: ~337.35 MB
  • Compressed Dataset Size: ~13.77 MB

Concluding Remarks

This investigation, forming a part of a broader exploration into data compression using various architectures, focuses specifically on the use of complex autoencoder structures. While simpler architectures could be employed, they typically result in greater information loss. My goal is to discover an architecture that not only compresses data but also does so in a near-lossless manner. The complex architecture of the autoencoders in this study was chosen to minimize information loss while achieving data compression. Moreover, the data compression analysis results highlight the efficiency of my autoencoder in terms of memory usage, making it suitable for applications requiring reduced data storage or bandwidth for data transmission. The ability of these models to reconstruct images serves as a testament to their potential for lossless data reuse. Moreover, the trained models have been saved and can be reused or fine-tuned for datasets with similar distributions, further enhancing their utility.

Acknowledgments

  • A special thank you to Dr. Peter Steinbach, Team Lead for Matter Research at Helmholtz AI Consultants, for his smart proposal regarding the topic of data compression using encoders. You can find more about his works on Peter Steinbach.
  • Data provided by MachineCurve.

Contact for Collaboration

  • If you are interested in collaborating or have ideas to share, please feel free to reach out to me at habibi.physics@gmail.com.

About

Exploring advanced autoencoder architectures for efficient data compression on EMNIST dataset, focusing on high-fidelity image reconstruction with minimal information loss. This project tests various encoder-decoder configurations to optimize performance metrics like MSE, SSIM, and PSNR, aiming to achieve near-lossless data compression.

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published