* Click the above links for details
- Design and development of multipurpose scale space filters
Design and development of multipurpose scale space filters
The detection of fine feature from an image is often challenging. Generally the smoothing filters used for the purpose are inadequate for such applications with the exception of bilateral and mean shift filter. A new scale space filter termed Gaussian Maxima Filter (GMF), developed using a parzen density estimation of a gaussian function shows considerably better performance in preserving fine features even for relatively large window size. The filter can be applied in different applications like noise removal, smoothing, image segmentation, image blending by altering the scale value.
- Object detection using a multi-parametric search space
Object detection using a multi-parametric search space
Consistent and reliable detection of features from an image is a challenging work. A new approach termed as Generalized Feature Vector, which essentially encapsulates and uses multiple features for consistent detection is proposed. The basic idea of GFV stems out from the fact that if detection relies on a single feature it may lead to multiple false alarms or wrong identification. This error can be greatly reduced to a large extent when a number of features are used instead of one, in an integrated framework. Though primarily applied to vision sensors GFV has the potential to include data obtained from other types of sensors. Experimental analysis of the proposed method shows that the approach is resilient to the extrinsic parametric variations with minimal false alarm. This further shows that the GFV method has tremendous potential in areas like robot navigation, surveillance, remote sensing and many more.
- Object tracking
One of the main challenges faced by object tracking and environment-modeling techniques is the frame-to-frame correspondence of the object of interest. False detections may lead to the tracking of wrong object thus misrepresenting information about the object location and its track. The tracking algorithm of the detected object should also be computationally inexpensive and suitable for real time applications. GFV, a multidimensional entity encapsulating multiple feature parameters, can uniquely identify dominant features of an object, and increase the detection reliability due to its potential to function consistently in any kind of environment, uninfluenced by view point invariance or extrinsic factors, thus generating minimal false alarms. Further a method to determine the 3D position of the object is presented which works on uncalibrated camera images and can be successfully applied to online processes. A statistical approach to reject outlier data, if any, is applied while identifying the trajectory.
- Visual Inspection of Braille Print quality
Visual Inspection of Braille Print quality
A major problem faced by the developing team is how to assess the quality of braille printout. In general a blind person is used to check the quality of the braille printed material.This by far is qualitative and often the assessment of the quality is subject to the reader's knowledge and experience in reading various printout produced by different printers.As a result it is felt that a vision based braille printout quality assessment system is essential. The GFV is used to extract the braille dots from the digital braille print, this is further utilized to measure the dot spacings for examining the printing quality. This helps in providing the necessary feedback to the design team so that they can improve the printer prototypes accordingly .A simple but elegant approach to convert the braille code to running text using single sided and interpoint Braille print is also proposed. The english text provided for printing and the reconverted english output can be compared to verify the accuracy of the braille coding system used during printing. Besides it can also be utilized in regenerating braille print from existing ones.
- Real time Panorama
Real time Panorama
A technique to generate a panoramic view in realtime by combining images acquired at different light exposures and camera orientations is proposed. Generalized Feature Vector (GFV) is used to detect features in order to find the corresponding feature points required to globally register the image and stitch them .A post processing operation over the composite image with a scale space filter called Gaussian Maxima Filter (GMF), removes the seamline and intensity discrepancies produced due to the varying light exposures of the individual images.
- Face recognition algorithms and related...
- Logarithmic Fourier PCA: A New Approach to Face Recognition
Logarithmic Fourier PCA: A New Approach to Face Recognition
This work proposes two face recognition algorithms namely Logarithmic Fourier Domain Principal Component Analysis (Log Fourier PCA) and Log fourier PCA with Independent Component Analysis (ICA) which successfully tackles multiple variations of face images. Neural network is used as classifier for both these methods. The Log Fourier PCA method proves to be resilient against illumination variations of the face images. However it is observed that the performance of this method decreases if there are multiple variations. Using ICA as a feature extractor further makes the recognition system robust to multiple parametric variations. Experimental results using Yale, ORL, FERET and PIE database shows that the proposed method Log fourier PCA followed by ICA and Neural network can tackle multiple variations in the face images. Their comparisons with other hybrid systems such as ICA-Neural Network and Fourier PCA-Neural network methods are also presented.
|Example face images from training set||Example images from test set, Green – Correct recognition, Red – Wrong recognition|
- Age Estimation using Gender Information
Age Estimation using Gender Information
Estimating age from a facial image is a intriguing and exigent task. Aging changes both shape as well as texture and it is an irreversible, uncontrollable and personalized. The way of aging in male is different from female and hence the accuracy of age estimation process can be improved if it is preceded by gender classification. The work proposed in this paper takes care of this by using gender information for categorizing age range of the given face image. Appearance parameters (AAM), containing shape and texture variations is used for gender classification which is analyzed with two well known classifiers Neural Networks and Support Vector machines (SVM). Gender classified appearance parameters are fed into male or female age estimator. Age estimation is then performed using Neural networks which classifies age range of the given face image. Experimental results on FG-NET age database demonstrate the effectiveness of the framework and validates that performance is better than existing approaches. The results also shows that appearance parameter from AAM increases the performance of the gender classification.
Age Estimation using Face images from FG-NET database
- Face and Periocular Biometrics
Face and Periocular Biometrics
This work presents a new multimodal biometric approach using face and periocular biometric. The available face recognition algorithm performance in presence of multiple variations such as illumination, pose, expression, occlusion and plastic surgery is not satisfactory. Also, periocular biometrics faces problem in presence of spectacles, head angle, hair and expression. The different conditions of the periocular and lips region along with the face image are shown in figure below. A method which can extract multiple feature information from a single source and can give a satisfactory performance even with less number of training images is desirable. Thus combining face and periocular data obtained from the same image may increase the performance of the recognition system. A detailed performance analysis of face recognition and periocular biometric using Gabor and LBP features is carried out. This is then compared with the proposed multimodal biometric feature extraction technique. The experimental results obtained using Muct and plastic surgery face database shows that the proposed multimodal biometric performs better than other face recognition and individual biometric methods.
Periocular and lips region along with the face image from Muct database
- Hand Gesture Recognition and related...
- Hybrid Approach for Sign Language Recognition
Hybrid Approach for Sign Language Recognition
Sign language communication used by hearing impaired people is a combination of letters (finger-spelled), words (mimicking the meaning), syllable (phonetics) and other gestural movements. These are the building blocks of sign language. The problem of sign language recognition can be divided into sub tasks like alphabet recognition, finger-spelled word recognition, gesture classification and continuous gesture recognition.
Two new approaches of hand gesture recognition which will recognize sign language gestures are developed. A hybrid feature descriptor, which combines the advantages of SURF & Hu Moment Invariant methods, is used as a combined feature set to achieve a good recognition rate along with a low time complexity. To further increase the recognition rate and make the recognition system resilient to view-point variations, the concept of derived features from the available feature set is introduced. K-Nearest Neighbour (KNN) and Support Vector Machine (SVM) are used for hybrid classification of single signed letter. In addition, finger spelled word recognition using Hidden Markov Model (HMM) for a lexicon based approach is also proposed.
- Indian Sign Language Recognition with global-local hand configuration
Indian Sign Language Recognition with global-local hand configuration
India is diversified in culture, language and religion. Since there is a large diversity among Indian languages, literature survey reports the non-existence of standard form of Indian Sign Language (ISL) gestures. ISL alphabets are derived from British Sign Language (BSL) and French Sign Language (FSL). Due to these issues, standard dataset for ISL alphabet/gestures have not been developed so far. Few research works has been carried out in ISL recognition and interpretation using image processing/vision techniques. But those are only initial work tried with simple image processing techniques and are not dealt with real time data.
Indian sign language uses both hands to represent each alphabet and gesture. The proposed approach addresses local-global configuration recognition, inter-class variability enhancement for each hand gesture. Hand region is segmented and detected by YCbCr skin color model reference.
|Indian Sign Language Alphabets
The shape, texture and finger features of each hand are extracted using Principle Curvature Based Region (PCBR) detector, Wavelet Packet Decomposition (WPD-2) and complexity defects algorithms respectively for hand posture recognition process. To classify each hand posture, multi class non linear support vector machines (SVM) is used. With this a recognition rate of 91.3% is achieved. Dynamic gestures are classified using Dynamic Time Warping (DTW) with the trajectory feature vector with 86.3% recognition rate.
- Hand Tracking and Hand-Face interaction gesture identification
Hand Tracking and Hand-Face interaction gesture identificatio
India is diversified in culture, language and religion. Since there is a large diversity among Indian languages, literature survey reports the non-existence ofThe problem of occlusion between hand and face is a biggest challenge put forth before the researchers. Moreover the interference between hand and face often occurs in continuous sign language conversation. Several body parts are involved in making a meaningful sign language gesture. For example in Indian Sign Language (ISL) face parts like chin, cheeks, lips, eyes, head, nose are referred to represent a gesture/sign.
|Sample signs of ISL with Hand-Face interaction
A robust approach to recognize hand gesture which involves face parts like chin, cheeks, eyes and head in the context of sign language recognition and also attacks the problem of interference between face and hand is developed. ICondensation algorithm is used to track the face, skin color segmentation is applied on face to eliminate eyes, and simple four quadrant multi clue information of face is obtained. Simultaneously two hands are tracked by another ICondensation module and BRIEF feature descriptors are extracted from hand. The multi clue information from four quadrants of face is identified whenever intersection of two tracking modules occurs. This intersection information provides which part of the face is being referred by the hand.
|Face tracking and Clue manipulation
Along with BRIEF feature descriptor, face position is used as feature for the gesture recognition. SVM multi-class classifier is used for continuous hand gestures classification. Experimentation is carried out with Indian Sign Language and found that the proposed approach outperforms the other existing methods with the recognition rate of 93.21%.
- Gesture Spotting for Continuous Sign Language Recognition
Gesture Spotting for Continuous Sign Language Recognition
The process of segmenting the start and end points of continuous gestures is called Gesture Spotting or Gesture Segmentation. This is an extremely difficult task due to the multitude of possible gesture variations in spatio-temporal space, shape in successive gestures and so on. Gestures may vary among person to person, time to time. Modeling the non-linear behavior of gestures with time is essential in case of continuous gestures recognition.
|Sample signs of ISL with Hand-Face interaction
The video is pre-processed and fed into the gesture spotting module. A two state Finite State Machine (FSM) will select the best prominent frames using the shape similarity and hausdroff distance measure. The selected frames are passed to grammar pruning section which will detect and reject the outlier gestures (like non gestural movements, transition gestures from one to another etc.,). The outlier gesture detection and rejection is performed with the reference model built using Hidden Markov Model (HMM) for both gesture and non gestural movement. As per this framework the video is now segmented into individual gestures. These individual segmented video is given to recognition module to map the corresponding hand signs. This gesture recognition module includes feature extraction which will handle some of the major issues of recognition, training and classification of gestures using well known classifiers like Dynamic Time Warping (DTW), Support Vector Machines (SVM) and Hidden Markov Model (HMM).