Abstract: In this paper, we present a direct solution method based neural network for image compression. The proposed technique includes steps to break down large images into smaller windows and eliminate redundant information. Furthermore, the technique employs a neural network that is trained by a non-iterative, direct solution method. An error backpropagation algorithm is also used to train the neural network, and both training algorithms are compared. The proposed technique has been implemented in C on the SP2 Supercomputer. A number of experiments have been conducted. The results obtained, such as compression ratio and transfer time of the compressed images are presented in this paper.
Introduction
In the late 1990s, the rapid growth of computer networks made the efficient transport of images a critical challenge. While traditional compression methods like predictive coding and vector quantization were established, Artificial Neural Networks (ANNs) emerged as a promising alternative due to their inherent ability to extract features and represent data in lower-dimensional spaces. However, the standard training algorithm, Error Backpropagation (EBP), is iterative and computationally intensive, often serving as a bottleneck for real-time applications. This paper explores a "Direct Solution Method" (DSM) to achieve efficient image compression with non-iterative training.
The Direct Solution Method (DSM)
The core innovation of this research is the DSM, which bypasses the slow gradient descent process of EBP. The authors utilize a Multi-Layer Perceptron (MLP) with a single hidden layer. By linearizing the output layer—specifically by applying the inverse of the sigmoid activation function—the weight-finding process is transformed into a linear system of equations.
As shown in the figure above, the activation function $f(net_k)$ is inverted to find $net_k$. This allows the weights to be solved directly using the Modified Gram-Schmidt method, significantly reducing training time compared to iterative methods.
Image Compression Pipeline
The proposed technique follows a multi-stage pipeline:
- Preprocessing & Binarization: Images are converted to monochrome using nearest-color thresholding to simplify the data.
- Segmentation: Images are divided into $8 \times 8$ or $16 \times 16$ pixel windows.
- Redundancy Removal: Identical or nearly identical windows are excluded from the training set to optimize efficiency.
- Compression: The network is trained such that the hidden layer acts as a bottleneck. For an $8 \times 8$ window (64 inputs), using 20 hidden units results in a compressed representation.
Results and Discussion
The system was tested on standard benchmarks like "Lena" and "Camera," as well as medical mammograms. Key findings include:
- Compression Ratio: A consistent 3:1 ratio was achieved (64 inputs to 20 hidden units).
- Transfer Efficiency: Compressed images were transferred via FTP in roughly 1/2 to 1/3 of the time required for original files. For instance, the transfer time for "Lena" dropped from 0.327s to 0.155s.
- Visual Fidelity: Reconstructed images showed minimal deterioration compared to the binarized originals.
- Security: The authors noted that the transmitted hidden layer outputs and weight matrices provide a basic level of inherent encryption, as the data does not resemble the original image during transit.
Conclusion
The Direct Solution Method offers a viable, high-speed alternative to traditional neural network training for image compression. While currently limited to binarized images and specific training sets, the approach demonstrates significant potential for reducing network load and improving data transfer speeds in specialized applications. Future work will focus on extending the method to grayscale images and improving generalization to unseen data.
Summary updated