PaperlayerView exact revision on arXiv

Source-linked AI summary

Indian Sign Language Recognition Using Eigen Value Weighted Euclidean Distance Based Classification Technique

Joyeeta Singha, Karen Das

arXiv:1303.0634v1cs.CV

TL;DR

Indian Sign Language recognition has seen few prior works, while interpreter access is described as difficult and unaffordable. The paper proposes a four-stage recognition system using eigenvalue-weighted Euclidean distance, reporting an improvement from 87% to 97%.

  • Problem

    The paper notes that few works address Indian Sign Language recognition and that finding experienced, qualified interpreters is difficult and unaffordable.

    Source: §I · p. 1

  • Method

    The system combines skin filtering, hand cropping, feature extraction, and classification using eigenvalue-weighted Euclidean distance between eigenvectors.

    Source: A. Skin Filtering · p. 3, B. Hand Cropping · p. 3, D. Classifier · p. 4

  • Results

    Success rate improved from 87% to 97% with eigenvalue-weighted Euclidean distance between eigenvectors as the classification technique.

    Source: §V · p. 4

  • Takeaways & Limitations

    The system recognizes static Indian Sign Language alphabets, including two-hand gestures, with a reported 97% success rate.

    Source: §VI · p. 7

    Evaluation used static images, excluding dynamic gestures H and J, and comprised only 240 images.

    Source: §VI · p. 7

Abstract

from arXiv · show

Sign Language Recognition is one of the most growing fields of research today. Many new techniques have been developed recently in these fields. Here in this paper, we have proposed a system using Eigen value weighted Euclidean distance as a classification technique for recognition of various Sign Languages of India. The system comprises of four parts: Skin Filtering, Hand Cropping, Feature Extraction and Classification. Twenty four signs were considered in this paper, each having ten samples, thus a total of two hundred forty images was considered for which recognition rate obtained was 97 percent.

I. INTRODUCTION

The paper addresses limited research on Indian Sign Language recognition and the difficulty and cost of finding interpreters. It reviews gesture-recognition approaches and proposes a system for Indian Sign Language alphabets.

I. INTRODUCTION

  • Few works had addressed Indian Sign Language recognition, while interpreters were described as difficult to find and unaffordable.The paper says computer translation of sign language to text could minimize the difference between deaf and hearing communities.Source: §I · p. 1
  • The paper proposes recognizing Indian Sign Language alphabets for human-computer interaction, with the stated aim of improving accuracy while minimizing time.The authors also suggest possible use in technology applications.Source: §I · p. 1

II. LITERATURE REVIEW

  • Prior gesture-recognition research spans segmentation, feature extraction, and classification, including skin filtering, gloves, PCA, neural networks, SVMs, and fuzzy logic.Reported limitations include single-hand recognition, slow processing, illumination sensitivity, and restricted gesture sets.Source: §II · pp. 1–2
  • The authors position their system as recognizing two-hand gestures with 97% accuracy using bare hands and low computational time.The review contrasts this with glove-based approaches and their sensor requirements.Source: §II · p. 2

A. Eigen value and Eigen vector

The passage describes eigenvectors as directions of variation in a covariance-based coordinate system and eigenvalues as measures of variance along those directions.

A. Eigen value and Eigen vector

  • Eigenvectors of the covariance matrix form an uncorrelated coordinate system, while their eigenvalues measure variance in that system.Selecting a few significant eigenvalues can reduce dimensionality for data compression.Source: A. Eigen value and Eigen vector · p. 2

IV. PROPOSED SYSTEM

The proposed system has four phases and uses a dataset of 240 camera-captured images representing 24 Indian Sign Language alphabets.

IV. PROPOSED SYSTEM

  • The system comprises skin filtering, hand cropping, feature extraction, and classification.The block diagram depicts the proposed system.Source: §IV · p. 2, Figure 1 · p. 2
  • The dataset contains 24 alphabets with 10 camera-captured samples each, totaling 240 images.Figure 2 shows some of the database images.Source: §IV · p. 2, Figure 2 · p. 3

A. Skin Filtering

Skin filtering separates skin-colored pixels from non-skin pixels, beginning with RGB-to-HSV conversion and ending by retaining the largest linked skin-colored region.

A. Skin Filtering

  • The filtering phase extracts skin-colored pixels from non-skin pixels in the input image.The paper presents this as the first phase of its system.Source: A. Skin Filtering · p. 3
  • RGB input is converted to HSV because the paper describes RGB as sensitive to illumination changes.The HSV model separates hue, saturation, and value.Source: A. Skin Filtering · p. 3
  • The HSV image is filtered and smoothened before conversion to a binary image of skin-colored pixels.The figure sequence shows RGB, HSV, filtered, smoothened, and grayscale binary images.Source: A. Skin Filtering · p. 3, Figure 5 · p. 3
  • The method retains the biggest binary linked object to exclude other skin-colored regions such as shadows, wood, or clothing.The retained region is described as the largest linked group of skin-colored pixels.Source: A. Skin Filtering · p. 3

B. Hand Cropping

Hand cropping keeps the hand portion through the wrist and clips away unnecessary image regions. The method scans the skin-filtered image to locate the wrist and crop the hand.

B. Hand Cropping

  • The crop retains the hand through the wrist while clipping away unnecessary image regions.Locating the wrist also helps locate the fingers, which lie on its opposite side.Source: B. Hand Cropping · p. 3
  • The skin-filtered image is scanned from multiple directions to detect the wrist and determine its position.The scan covers left, right, top, and bottom.Source: B. Hand Cropping · p. 4
  • The image is cropped using the detected wrist position and the minimum and maximum white-pixel coordinates.Figure 6 shows input images alongside their cropped versions.Source: B. Hand Cropping · p. 4, Figure 6 · p. 4

C. Feature Extraction

Feature extraction computes eigenvalues and eigenvectors from the cropped image. The system uses the five most significant eigenvectors, reducing the feature representation from 50 dimensions.

C. Feature Extraction

  • The cropped image is the input for computing eigenvalues and eigenvectors.The input image is specified as 50 by 50 pixels.Source: C. Feature Extraction · p. 4
  • The system computes the covariance matrix and derives eigenvectors and eigenvalues from it.Eigenvectors are ordered by decreasing corresponding eigenvalue.Source: C. Feature Extraction · p. 4
  • Five significant eigenvectors out of 50 are retained because the remaining eigenvalues are very small.The authors describe this as reducing dimensionality and compressing data without much information loss.Source: C. Feature Extraction · p. 4

D. Classifier

The classifier uses two levels: Euclidean distances between test and database eigenvectors, followed by eigenvalue-weighted distances to select a recognized symbol.

D. Classifier

  • The proposed classifier uses eigenvalue-weighted Euclidean distance between eigenvectors in a two-level classification process.The second level weights the first-level Euclidean distances by differences between test-image and database-image eigenvalues, sums the results for each image, and selects the minimum.Source: D. Classifier · p. 4
  • At the first level, five Euclidean distances are computed between test-image and database-image eigenvectors, and the minimum is found for each database image.EV1 denotes test-image eigenvectors and EV2 database-image eigenvectors.Source: D. Classifier · p. 4

V. RESULTS AND DISCUSSIONS

The two-level classification comparison reports a success-rate increase from 87% to 97% with eigenvalue-weighted Euclidean distance. The system also reports 97% recognition and an image-processing time of 0.0384 seconds.

V. RESULTS AND DISCUSSIONS

  • 97% recognition rate was obtained in tests of different images using the new classification technique.Tables I and II show examples of the two classification levels on test images.Source: §V · p. 4, Table I · p. 5, Table II · p. 6
  • 87% to 97%: the success rate improved with eigenvalue-weighted Euclidean distance between eigenvectors.Table III presents success rates for the two classification levels.Source: §V · p. 4, Table III · p. 7
  • 0.0384 seconds was the reported time to process an image.The authors report recognition of different Indian Sign Language alphabets.Source: §V · p. 7

VI. CONCLUSION AND FUTURE WORK

The system uses eigenvalue-weighted Euclidean distance to recognize Indian Sign Language alphabets and reports a 97% success rate. Its evaluation is limited to static images, excluding H and J, and the authors propose expanding the dataset and addressing dynamic gestures.

VI. CONCLUSION AND FUTURE WORK

  • 97% success rate was achieved using eigenvalue-weighted Euclidean distance to recognize Indian Sign Language alphabets.The system considered both hands.Source: §VI · p. 7
  • Dynamic gestures H and J were excluded because experiments used static images; the authors propose addressing dynamic gestures in future work.The authors also hope to extend the 240-image dataset.Source: §VI · p. 7
Loading 1303.0634v1…