RESEARCH

Image & Vision

Investigates Image data fusion techniques that combine image and track data from multiple sensors to achieve improved accuracies and more specific inferences than could be achieved by the use of a single sensor alone. Our aim is to explore the state-of-the-art image processing algorithms for achieving effective data fusion as in:

Classification and Recognition

작성자 관리자 날짜 2021-04-08 14:51:40 조회수 39

1. Gradient-based Feature Extraction        

    Feature Extraction technique is important to effective presentation of image information and accuracy of recoginition. The features are extracted from gradient vectors that are not much affected by illumination. Following demostration is the implementation of  the feature extraction.

2. Pose-invariant face recognition        

   The idea here is to transform a face image to front image using estimated angle information. On the assumption that face is shaped like that of a cylinder, we estimate the object's yaw pose and then extract the frontal face image via a yaw pose transform with previously estimated yaw pose angle. In addition, a pitch pose is also projected to the frontal face to capture the pitch variations and thus increase the overall recognition performance under yaw and pitch motions of object face. A stereo camera is employed to estimate the object's pitch pose and then project the object's pitch pose to that of frontal face, similar to that of pose projection. The projected and normalized features are used to increase the overall recognition rate by decision-level fusion between two images of stereo camera.

 

3. Fingerprint Recognition       

   Fingerprint recognition consists of reference point detection part and feature extraction part. First, reference point is detected by exploiting the gradient probabilistic model that captures the curvature information of fingerprint texture. The reference point detection is accomplished through searching and locating the points of occurrence of the most evenly distributed gradient in a probabilistic sense. The uniformly distributed gradient texture represents either the core point itself or those of similar points that can be used to establish the rigid reference from which to map the feature for recognition. Once a reference point is established, the next step in the fingerprint recognition process is to extract unique and characterizing features of the fingerprint to be able to distinguish it from others. For feature extractions, we use the fingercode extraction technique based on a filterbank. This method is the most reliable approach due to its scale, translation and rotation invariant properties. Here, features are extracted by using mean and variance of interest region from reference point. The next avi file shows an example of fingerprint recognition.

  • Fingerprint recognition based on a filterbank

4. A number plate Recognition        

   A number plate recognition consists of alignment part, segmentation, feature extraction and recognition part. After segmentation of a number plate, we use neural network for recognition. The following avi file shows a example of a number plate recognition procedure.

  • A number plate recognition using neural network

5. Human and Vehicle Recognition        

  • Classification between human and vehicle.
  • Classification step

                  - Detection of moving object
                  - Feature extraction of detected object
                  - Recognition using neural network

 

6. Human and Vehicle Recognition using star skeleton        

  • Classification between human and vehicle based on the star skeleton feature

                - Human : non-rigid characteristic
                - Vehicle : rigid characteristic

  • Classification step

                - Detection of moving object
                - Finding the center point and five far-point in detected regions
                - Classification using variation of points (rigid or non-rigid)

7. Text Recognition for broadcast news        

   This demo is "Extraction and Recognition of Korean Text in Broadcast news". The Region of text is extracted by using high frequency component and edge information in color domain. This region is converted to binary domain using dynamic threshoding method. after that, dividing of region into each character is perfomed. Neural network is applied to classify character into 6 types according to shape And, after divide character into vowel and consonant, recognizing the character is performed using Neural network.

댓글 (0)

등록된 댓글이 없습니다.
작성 권한이 없습니다.