|Title:||An autoregressive model approach to object recognition and its application to handwritten numeric character recognition|
|Subject:||Optical pattern recognition|
Image processing -- Digital techniques
Writing -- Data processing
Hong Kong Polytechnic University -- Dissertations
|Pages:||vii, 96 leaves : ill. ; 30 cm|
|Abstract:||Digital image processing has been rapidly developing in different areas, e.g. medical image processing such as tomography and thermography; earth resource inventory such as land usage and minerals; office automation such as document storage, retrieval and reproduction; industrial production such as computer vision for mechanical robots. These different applications all involve object or pattern recognition which is extensively using in the sectors of industrial production and commercial products. The ultimate aim of these image processing applications is to extract important features from a picture or an image data, from which a description, interpretation, or understanding of the scene can be provided by the machine. More sophisticated vision systems are able to describe the various objects and their relationships in the scene. In general, an object recognition system contains Image Enhancement Unit, Feature Extraction Unit and Feature Classification Unit. A picture is input to the object recognition system with the description of the object inside the picture as the output. The picture is enhanced in the Image Enhancement Unit, then the object is segmented and its features are extracted in the Feature Extraction Unit. Lastly, the features extracted are used to classify the object to the corresponding object class in the Feature Classification Unit. In this dissertation, we will discuss each subprocess in a typical image processing system in details. A conceptual introduction to object recognition is given in Chapter 1 and the Image Segmentation will be discussed in Chapter 2. In Chapter 2, types of feature that can be extracted from an image is listed. We will based on a boundary-approach to segment an object from its background by using a technique called "Boundary Follower". The algorithm for this boundary follower will be explained and illustrated. To complete the study of various segmentation methods, we will discuss various methods e.g. Chain Codes, Fitting Line Segments, B-Spline Representation, Fourier Descriptors, Autoregressive Model to represent an object boundary in Chapter 3. Among the various boundary representation methods, we chose the Autoregressive (AR) Model to represent an object boundary. In Chapter 4, we will discuss various type of AR Models and the properties of each type of AR Models and the consideration for choosing the model order used in the AR Model. In our experiment, we used one-dimensional AR Model using equispaced radius vector as our feature extraction model. A time series which is used to represent a boundary object will be transformed to a set of AR parameters and these AR parameters form a feature vector which is used to represent a specific boundary shape. Generally speaking, the time series used in this 1-D AR model is good for convex or wisesense convex object boundary. We will discuss an "unwrapping" technique to handle those non-convex object boundaries. In addition, the technique to solve the "line segment" problem is also discussed. Lastly, the algorithm for the boundary approximation by a time series is also included in Chapter 4. The feature vectors extracted by AR Model can then be used for classification in the Feature Classification subprocess which is discussed in Chapter 5. Various classification methods are discussed including both Supervised Learning Classifiers and Nonsupervised Learning Classifiers. The supervised learning classifier such as Minimum Distance Classifier, Decision Tree Classifier, Bayes' Minimum-risk Classifier and the nonsupervised Learning such as Chain method, Isodata, Feature Weighing Method are explained. An experiment on handwritten numeric character was conducted using an 1-D AR Model as the Feature Extraction Model and a number of classifiers were used in the Feature Classification subprocess. The experiment is summarized in Chapter 6 which includes the "Training Stage" and "Recognition Stage". 50 training image samples were used in the training stage to calculate the information used in the recognition stage. These information includes the mean feature vector for each class (i.e. digit "0" to '9') and the standard derivation of the each feature for each class. A decision tree was also built as a decision tree classifier in the training stage. In the recognition stage, 14 testing images were used to test the ability of the AR parameters for object recognition. These testing images include both scaled, rotated and occluded numeric character image to test the inherent invariant properties of the AR Model. In the experiment we also tested the effect of the shifted object centroid on the object recognition results. Lastly, we will discuss various factors affecting the classification accuracy and their modification in the conclusion.|
Files in This Item:
|b11834365.pdf||For PolyU Staff & Students||2.76 MB||Adobe PDF||View/Open|
As a bona fide Library user, I declare that:
- I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
- I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
- I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.
By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.
Please use this identifier to cite or link to this item: