Efficient Noise Reduction and HOG Feature Extraction for Sign Language Recognition
MATLAB implementation of the research paper βEfficient Noise Reduction and HOG Feature Extraction for Sign Language Recognitionβ, presented at the 2018 International Conference on Advancement in Electrical and Electronic Engineering (ICAEEE 2018).
The project implements an image-based American Sign Language (ASL) alphabet recognition system using image preprocessing, hand segmentation, Canny edge detection, Histogram of Oriented Gradients (HOG) feature extraction, and a K-Nearest Neighbors (KNN) classifier.
The proposed approach achieved 94.23% classification accuracy on the reported test set, outperforming the Bag-of-Features + SVM comparison method, which achieved 86%.
π Research Paper
Efficient Noise Reduction and HOG Feature Extraction for Sign Language Recognition
Authors
- Iqbal Mahmud
- Tasnim Tabassum
- Md. Palash Uddin
- Emran Ali
- Adiba Mahjabin Nitu
- Masud Ibn Afjal
Published at ICAEEE 2018, Gazipur, Bangladesh.
DOI: 10.1109/ICAEEE.2018.8642983
π Overview
Automatic sign language recognition is a computer vision problem where visual hand signs are classified into their corresponding characters.
This project follows the approach proposed in the paper:
Input Image
β
βΌ
Logarithmic Transformation
β
βΌ
Histogram Equalization
β
βΌ
L*a*b Color-Space Segmentation
β
βΌ
Binary Image
β
βΌ
Canny Edge Detection
β
βΌ
HOG Feature Extraction
β
βΌ
KNN Classifier
β
βΌ
Recognized ASL Character
The preprocessing stage is designed to reduce background noise and isolate the hand before extracting shape-based HOG features.
π§ Method
1. Image Preprocessing
The input images are enhanced using image-processing operations described in the paper.
Logarithmic Transformation
A logarithmic transformation is applied to improve image intensity distribution and enhance relevant visual information.
Histogram Equalization
Histogram equalization is then used to improve contrast and stretch the available intensity range.
2. Hand Segmentation
The hand is separated from the background using Lab color-space segmentation.
The repository includes an implementation generated using MATLAB's Image Segmenter workflow in:
This produces a binary/masked representation of the hand region, reducing the influence of background objects.
3. Canny Edge Detection
After segmentation, the image is converted into a binary representation and edge information is extracted using the Canny edge detector.
The resulting edges provide the structural information needed for shape-based feature extraction.
4. HOG Feature Extraction
The proposed method uses Histogram of Oriented Gradients (HOG) to describe the shape and edge structure of each sign.
The main implementation is:
ASLRecognitionWithHOGFeaturesAndKNNClassifier.m
The implementation uses MATLAB's extractHOGFeatures() function and stores the resulting feature vectors in a training feature matrix.
5. KNN Classification
The extracted HOG features are classified using a multi-class K-Nearest Neighbors (KNN) classifier.
The classifier is created using MATLAB's:
The system recognizes all 26 English alphabet characters (AβZ).
π Dataset
The experiment described in the paper uses an American Sign Language alphabet dataset containing:
- 26 classes: AβZ
- 572 total images
- 22 images per character
- 200 Γ 200 Γ 3 RGB images
- 65% training / 35% testing split
- 14 training images per character
- 8 testing images per character
The repository expects the image dataset to be organized into MATLAB imageSet-compatible class directories.
Note: The dataset itself is not included in this repository.
π Results
The paper reports the following overall result on the test set:
| Method |
Classifier |
Test Accuracy |
| Proposed Method |
HOG + KNN |
94.23% |
| Comparison Method |
Bag of Features + SVM |
86.00% |
The proposed method correctly classified 196 out of 208 test samples.
Per-class results
| Character |
Accuracy |
| A |
100% |
| B |
100% |
| C |
100% |
| D |
100% |
| E |
100% |
| F |
100% |
| G |
100% |
| H |
100% |
| I |
100% |
| J |
100% |
| K |
100% |
| L |
87.5% |
| M |
100% |
| N |
100% |
| O |
75% |
| P |
100% |
| Q |
100% |
| R |
100% |
| S |
50% |
| T |
100% |
| U |
100% |
| V |
75% |
| W |
100% |
| X |
87.5% |
| Y |
100% |
| Z |
75% |
π Comparison Method
For comparison, the project also includes a Bag-of-Features implementation:
ASLRecognitionWithBagOfFeatures.m
This approach uses MATLAB's Bag-of-Features pipeline and an image category classifier.
The implementation:
- Loads the training and test sets
- Builds a Bag-of-Features representation
- Trains an image category classifier
- Predicts the test set
- Generates a confusion matrix
The paper reports 86% test accuracy for this comparison approach.
π Repository Structure
.
βββ TrainingMatrix/
β
βββ ASLRecognitionWithHOGFeaturesAndKNNClassifier.m
βββ ASLRecognitionWithBagOfFeatures.m
βββ GenarateConfusionMatrix.m
βββ PlotConfusion.m
βββ SegmentImage.m
βββ batchmyimfcn.m
βββ myimfcn.m
β
βββ LICENSE
Main files
| File |
Description |
ASLRecognitionWithHOGFeaturesAndKNNClassifier.m |
Main HOG feature extraction and KNN classification pipeline |
ASLRecognitionWithBagOfFeatures.m |
Bag-of-Features comparison implementation |
SegmentImage.m |
Hand/background segmentation function |
GenarateConfusionMatrix.m |
Generates classification confusion matrix |
PlotConfusion.m |
Visualization of confusion matrix |
myimfcn.m |
Image processing function |
batchmyimfcn.m |
Batch image processing |
TrainingMatrix/ |
Training-related data/matrices |
π οΈ Requirements
The implementation is written in MATLAB and uses functionality from MATLAB's computer vision/image processing ecosystem.
Recommended:
- MATLAB
- Image Processing Toolbox
- Computer Vision Toolbox
- Classification / Statistics and Machine Learning functionality
The exact toolbox requirements may vary depending on the MATLAB release being used.
π Running the Project
1. Clone the repository
git clone https://github.com/zero-jim/Efficient-Noise-Reduction-and-HOG-Feature-Extraction-for-Sign-Language-Recognition.git
cd Efficient-Noise-Reduction-and-HOG-Feature-Extraction-for-Sign-Language-Recognition
2. Prepare the dataset
Place the ASL dataset in the expected directory structure.
The main HOG implementation loads the dataset using:
handDatabase = imageSet('DataATT','recursive');
Therefore, the dataset should be available under:
DataATT/
βββ A/
βββ B/
βββ C/
βββ ...
βββ Z/
Each directory should contain the corresponding character images.
3. Run the HOG + KNN implementation
Open:
ASLRecognitionWithHOGFeaturesAndKNNClassifier.m
and run the script from MATLAB.
The script:
- Creates the image database
- Splits the data into training and testing sets
- Extracts HOG features
- Trains the KNN classifier
- Predicts test images
- Displays the recognized sign
The repository implementation uses a 65:35 training/testing split, consistent with the experimental setup described in the paper.
π§ͺ Reproducing the Paper
To reproduce the main experiment:
- Obtain the ASL alphabet dataset used in the paper.
- Arrange the images into the expected class-directory structure.
- Run the preprocessing/segmentation pipeline.
- Run
ASLRecognitionWithHOGFeaturesAndKNNClassifier.m.
- Generate the confusion matrix using the provided scripts.
- Compare the result with
ASLRecognitionWithBagOfFeatures.m.
The published experiment used 364 training samples and 208 testing samples, with 196 of the 208 test samples correctly recognized by the proposed method.
π Citation
If you use this implementation or the underlying research, please cite the original paper:
@inproceedings{mahmud2018efficient,
title={Efficient Noise Reduction and HOG Feature Extraction for Sign Language Recognition},
author={Mahmud, Iqbal and Tabassum, Tasnim and Uddin, Md. Palash and Ali, Emran and Nitu, Adiba Mahjabin and Afjal, Masud Ibn},
booktitle={2018 International Conference on Advancement in Electrical and Electronic Engineering (ICAEEE)},
year={2018},
doi={10.1109/ICAEEE.2018.8642983}
}
π License
This repository is released under the Apache License 2.0.
See LICENSE for the complete license text.
π References
π Acknowledgment
This repository contains the code associated with the research paper βEfficient Noise Reduction and HOG Feature Extraction for Sign Language Recognition.β
The implementation is intended to facilitate reproducibility and provide a practical reference for classical computer-vision-based sign language recognition using image preprocessing, HOG features, and KNN classification.