The Research of Image Information Processing Based on the Improved BP Algorithm

Feng ZhaoThe Open Automation and Control Systems Journal 2015 · DOI

Disclaimer: This article summarizes a retracted research paper. The following content is presented for informational purposes regarding the proposed methodologies and reported results.

The Research of Image Information Processing Based on the Improved BP Algorithm

Abstract

Traditional Backpropagation (BP) algorithms often suffer from slow convergence and susceptibility to local optima. This paper presents an improved BP algorithm incorporating an adaptive learning rate and momentum factors. Applied to image fusion and filtering, the algorithm effectively reduces noise while preserving edge sharpness, overcoming the blurring common in traditional filters. Experimental results demonstrate that the improved BP network significantly accelerates training and enhances accuracy in complex tasks such as fingerprint identification.

Introduction

In image processing, a fundamental challenge exists in the trade-off between noise reduction and detail preservation. While traditional spatial filters can mitigate noise, they often introduce blurriness that obscures critical features like edges. Neural networks, particularly BP networks, offer a non-linear mapping capability suitable for image fusion and denoising. However, the standard BP algorithm's efficiency is limited by slow convergence and local minima traps. This research proposes algorithmic improvements to create a faster, more reliable BP network for high-stakes computer vision tasks.

Methodology

The proposed Improved BP Neural Network utilizes a standard three-layer architecture consisting of input, hidden, and output layers.

Standard three-layer BP neural network architecture featuring input, hidden, and output layers.

To address the inherent weaknesses of the standard BP algorithm, two primary modifications are introduced:

  1. Adaptive Learning Rate: Instead of a static value, the learning rate is dynamically adjusted. It begins with larger steps to accelerate initial learning and decreases over time to ensure stable convergence near the target minimum.
  2. Momentum Factor: By incorporating a portion of the previous weight update into the current iteration, the network gains "momentum." This helps the training process bypass local optima and smooths the overall learning trajectory.

Experimental Applications

The improved algorithm was evaluated across two distinct domains:

Image Fusion and Denoising

The network was trained using images corrupted by varying levels of Gaussian noise. Unlike traditional filters that apply a spatial blur, the neural network learns the underlying visual patterns to fuse multiple noisy instances into a single, clean output. This method aims to preserve sharp edges while mathematically smoothing out noise.

Fingerprint Identification

The authors applied the improved BP network to a database of 3,306 fingerprint images. The system extracts features based on the directional field of ridges and compares single-feature matching against fusion-based methods that integrate multiple extraction techniques, such as minutiae, texture, and pattern recognition.

Key Results

The improved BP algorithm demonstrated a significant performance boost in both speed and accuracy:

  • Convergence Speed: In image fusion tests, the training time required to reach a target error of 0.06 was reduced from 24,501 ms (standard BP) to 4,436 ms—an approximately 82% reduction in training time.
  • Recognition Accuracy: In fingerprint identification, the fusion-based approach achieved a first-candidate recognition rate of 98.73%, consistently outperforming single-feature methods (which ranged from 83.68% to 98.23%).
  • Efficiency Trade-offs: While the highly accurate fusion methods required slightly more processing time (~0.6s per identification) compared to single-feature methods (~0.1s to 0.28s), the substantial gain in reliability justifies the computational cost for security-sensitive applications.

Conclusion

The improved BP algorithm effectively addresses the convergence and local optima issues characteristic of standard neural networks. By applying these optimizations to image fusion and biometric identification, the research demonstrates superior noise reduction and recognition accuracy compared to traditional methodologies. The results suggest that ensemble-based feature fusion is a robust strategy for enhancing the performance of automated identification systems.

Summary updated

Abstract

As the original BP algorithm has slower convergence speed and the training process is easy to be trapped in local optimal solution. In the paper we put forward an improved algorithm and put it apply in the traditional filter. It solves the problem that when the traditional filter reduces the noise it will become fuzzy. And the algorithm reaches the purpose of image fusion and filtering the image noise. Finally, the experiment shows that the image fusion method based on improved BP neural network is superior to the traditional filter method. It can reduce the noise effectively and further it indicates that the performance of fusion-based method is better than any method used alone in fusion and can be used in fingerprint identification.
BibTeX
@article{zhao2015research,
  author = {Feng Zhao},
  title = {{The Research of Image Information Processing Based on the Improved BP Algorithm}},
  journal = {The Open Automation and Control Systems Journal},
  year = {2015},
  doi = {10.2174/1874444301507011112},
}
Figure