A Hybrid Learning Approach for Better Recognition of Visual Objects

AAAI 1996

TL;DR

This paper introduces a novel two-phase approach to improve visual object recognition, especially in the presence of variations like noise and rotation. By first learning individual classifiers for altered object versions and then intelligently structuring these classifiers using a decision tree learning system, they achieve significantly better recognition rates than traditional single-classifier or ensemble methods.

What Problem Does This Paper Address?

Recognizing objects in real-world images is challenging because objects can appear with variations in rotation, noise, lighting, and other visual alterations. Traditional computer vision systems often struggle with these variations, leading to lower recognition accuracy. This paper aims to develop a more robust system that can reliably identify objects even when they are presented in non-ideal conditions. The core problem is how to effectively combine the knowledge from multiple specialized classifiers, each potentially tuned to a specific type of visual alteration, into a single, high-performing recognition system.

Methods

The proposed approach is a two-phase hybrid learning system:

  • Phase I: Learning Classifications of Visual Objects:
    • Data Augmentation: For each standard object image, multiple "altered" versions are created by applying transformations like adding Gaussian noise and geometric rotations.
  • Specialized Classifiers: For each type of alteration (and different parameter settings), a separate classifier (specifically, Radial Basis Function - RBF networks are used) is trained. This results in a set of classifiers, each potentially good at recognizing objects under specific conditions.
  • Record of Recognition: For each image, the outputs from all trained classifiers are collected along with the correct classification. This forms a "record of recognition."
  • Phase II: Learning Optimized Structures of Classifiers:
    • Decision Tree Learning: The collected "records of recognition" are used as training data for a decision tree learning system (AQDT-2).
  • Optimized Structure: AQDT-2 learns an optimal decision structure (a tree) where each node represents a classifier. This structure is designed to minimize the number of classifiers used and maximize recognition accuracy. The system intelligently decides which classifier to use at each step, effectively creating a "plan" for recognition.

Datasets

  • Hand Gesture Database: The authors created a custom database of 15 hand gestures. Images were taken from 5 different individuals, with two sets of images per person, one taken after a 30-minute gap. For experimental complexity, subsets of gestures were grouped into single classes.
  • Data Augmentation: The original images were augmented by:
  • Adding Gaussian noise (mean 0, variance 10).
  • Applying a geometric transformation (10-degree rotation).
  • Training/Testing Split: The data was split into training and testing sets. For the ERBF (Ensemble of RBF) and AQDT-2 phases, specific splits were used, with training set #1 for training classifiers and training set #2 for testing. The paper also mentions using a 9-cycle, two cross-fold testing method.

Key Insights and Ideas

  • Specialization and Generalization: The core idea is that no single classifier can handle all variations. By training specialized classifiers for different types of alterations and then learning how to best combine them, the system gains both specialization and generalization capabilities.
  • Intelligent Decision Making: The AQDT-2 system acts as an intelligent orchestrator, learning when to use which specialized classifier, rather than just averaging their outputs (as in a simple ensemble). This leads to a more efficient and accurate decision process.
  • Structure Matters: The paper emphasizes that the structure of how classifiers are combined is crucial. A well-learned decision tree can significantly outperform a flat ensemble of classifiers.
  • Complexity as a Feature: Grouping gestures into single classes to increase problem complexity was a deliberate choice to better test the robustness of the proposed method.

Key Results

The paper reports significant improvements in recognition rates compared to using a single RBF classifier or an ensemble of RBF classifiers (ERBF).

  • ERBF vs. Single RBF: Tables 1 and 2 show that the ERBF method (combining 9 classifiers) consistently outperforms a single RBF classifier across different testing combinations, with accepted (correct) rates generally higher and rejected (false positive/negative) rates lower.
  • Hybrid Approach (ERBF + AQDT-2) vs. Others: Table 3 and Figure 7 demonstrate the superiority of the proposed hybrid approach.
  • The error rate for the hybrid approach averaged 3.05% across testing combinations.
  • This is substantially lower than the ERBF (majority voting) which averaged 3.1% (though the paper states ERBF is better than single RBF, the average error rate is very close, suggesting AQDT-2 provides the key improvement).
  • The single RBF classifier had significantly higher error rates (e.g., 27.2% to 43.1% for false negative in Table 1).
  • Optimized Structure Efficiency: Figure 6 shows an example optimized decision structure that uses an average of 3 classifiers to classify a gesture, with some gestures being classified by as few as 4 networks.

Limitations and Future Work

  • Limited Alterations: The study only considered two types of visual alterations: Gaussian noise and geometric rotation. The authors suggest that more complex alterations and a wider variety of transformations should be explored.
  • Number of Classifiers: The paper uses a fixed number of 9 classifiers (3 RBF components, each with 3 nodes). Future work could investigate how the number of these base classifiers affects the final performance and structure complexity.
  • Dataset Size and Diversity: While the hand gesture dataset was augmented, it's a relatively small and specific domain. Applying the approach to larger, more diverse datasets would be valuable.
  • Analysis of Structure Complexity: Further analysis is needed to understand the relationship between the number of alterations used, the complexity of the learned decision structures, and the overall object recognition performance.

Summary updated

BibTeX
@inproceedings{1996hybrid,
  title = {{A Hybrid Learning Approach for Better Recognition of Visual Objects}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1996},
}

Loading related content...

Figure