Dissertation > Excellent graduate degree dissertation topics show

Learning Simple Local Features for Object Detection

Author: ZhangWeiZe
Tutor: DongJinXiang;TongRuoFeng
School: Zhejiang University
Course: Computer Science and Technology
Keywords: Object Detection Computer Vision Changes within the class Variation between classes Local Feature Haar-like features Template Integral image Hough Transform Implicit Shape Model Edge Detection Weak hypothesis Strong assumptions Maximum and contiguous subsequence Dynamic Programming Testing standards
CLC: TP391.41
Type: PhD thesis
Year: 2010
Downloads: 236
Quote: 0
Read: Download Dissertation

Abstract


This paper studies the field of computer vision, object detection is a very challenging task, from real-world images to identify certain types of objects and images in the background chaos locate the object. This is the human visual system is one of the core competencies, but the computer vision system is still far from close to the level of the effect. The issue is the major difficulty is finding an efficient object representation, one must tolerate similar objects of different classes varies between individuals, on the other hand be able to distinguish between different types of objects in the class change, but also need a robust processing confusion background illumination changes and partial occlusion and other issues. This paper reviews the research status object detection methods, analyzes and summarizes the common object model and features, from simple local feature and feature-based learning algorithms two aspects of how to build an accurate and robust object model. The main research results and innovations are the following two aspects: the feature level, this paper presents a table-based form of local features - scattered rectangle feature, and one based on the shape of local features - Hough transform for line segment (group ). Dispersing the class rectangle features a variant of Haar features. And Haar-like features, as it is a template-based simple rectangular features, but a rectangular template in the horizontal or vertical direction without adjacent alignment, and therefore not only an arbitrary orientation information can be expressed, and can be expressed on the geometry shear, separation and overlapping shape information, characterized in that the rectangular distributed representation of the object to be more flexible section having a better representation capability. Meanwhile, you can use integral image, images calculated in constant time window of any size anywhere on the eigenvalues ??of the dispersion rectangle, adjacent to solve the lack of alignment constraints brought by the large number of features of computational problems. In addition, the use of construction methods, strict proof of any non-degenerate rectangle dispersion characteristics and satisfy certain geometric constraints between multiple classes equivalence between Haar features. The description of a non-degenerate Equivalence rectangle feature of the dispersion pieces of information included in the object class of the plurality of Haar features equivalent to the information contained in the integrated, so the more robust features. At MIT and cMU face test set comparative experiments showed scattered rectangle features based classifier outperforms Haar-like feature classifiers. Hough transform line (group) is made by the line drawing inspired a simple shape feature. This feature is not that of its two endpoints, but by which method the angle between the abscissa, the vertical distance between the origin of coordinates, the vertical distance from the center line and the line segment length consisting of the four-tuple. The four-tuple representation not only uniquely determine an arbitrary line, and can easily handle scaling, rotation and translation transformations. Center of the object given by the local coordinate system as the origin, the Hough transform the center line and the geometric relationship between the object to be implied in the tuple in a compact implicit shape model, which has been the object Detection proved to be an effective model. Hough transform constitutes a connected line segments Hough transform group, segment by introducing local geometric information between features to further enhance the ability to distinguish between. Hough transform line segment (group) the degree of similarity between the representation space by a four-tuple of a weighted Euclidean distance measure. Quad element by adjusting the corresponding weight, the distance can better tolerate a reliable edge detection caused by noise, which can be used to select the sample from the training of the Hough transform with ability to distinguish between line segment group, the establishment of the class object code form. Achieved by the shape matching the shape of the object detection experiments show indeed an important feature defined in the object class, object detection task can be qualified. In the learning algorithm level, this paper presents a variant of AdaBoost algorithm - Double Threshold AdaBoost algorithm. The variant with the original algorithm using the same framework, but with two thresholds using weak assumptions, the core idea is to choose a better classification performance weak hypothesis can make learning to assume a more robust and efficient strong. Dual-threshold weak hypothesis subside into the weak hypothesis of a single threshold case, the characteristic value of the sample space is more fine-grained, ensures that the classification error weak hypothesis than a single threshold value is smaller. In order to quickly determine the optimal value of the two thresholds, the corresponding paper proposes a dual-threshold weak learning algorithm. The weak learning algorithm will determine the threshold question is converted to find the maximum and continuous subsequence problem, which can be used to solve linear dynamic programming algorithm. Haar features and dispersing the class rectangle feature learning results show that the performance requirements of the same training, the speed of convergence of the variant, the resulting detector has fewer levels, with fewer features. CMU face the MIT test set and the test results show that the variant learning performance of the overall classification of the AdaBoost algorithm is better than the original. In addition to the above results, the test article in the face detection process, to build a 19 × 19 pixel resolution of the front face training set, presents a face against MIT and CMU test set testing standards. The standard use of the information provided on the test set into the facial features of the real location information into the face rectangle defines the minimum and maximum possible distribution of correct detection area can be used as an objective and rigorous testing standards.

Related Dissertations

  1. Preparation of ITO Quai-1D Nanostructures by Sol-Gel Method Combined with Porous Anodic Aluminum Oxide Template,TB383.1
  2. Study on the Synthesis Bi3.25La0.75Ti3O12(BLT) Nanotubes and Nano Wires,TB383.1
  3. Hydrothermal Synthesis of Oxide Hollow Spheres,TB383.4
  4. Research on Algorithms of 2D Face Template Protection,TP391.41
  5. Research on Learning-Based Low-Level Vision Problem,TP391.41
  6. The Research of Moving Object Tracking System Based on Embeded Image Process Unit,TP391.41
  7. Research on Visual Detection and Tracking of Mobile Robots,TP242.62
  8. Research on Secure Fingerprint Authentication Based on Distance Projection Coding,TP391.4
  9. Research on Nondestructive Detection Technology for External Qualities of Papayas Based-on Vision,S667.9
  10. Research on Detecting Optical Fiber Geometric Parameter Based on Machine Vision,TN253
  11. Research on Recognition of Fabric Weave Pattern Based on Space-Frequency Domain,TS101.923
  12. Study on Detection and Grading of ’Jiro’ Persimmon’s External Quality Based on Computer Vision,S665.2
  13. Prediction of Pork Processing Functionalities,TS251.1
  14. Research on Inspection Technology of Dehydrated Garlic Slice Based on Computer Vision,TP391.41
  15. Incomplete information on the completeness of the system and its knowledge acquisition,TP311.13
  16. Moving target trajectory analysis based Intelligent Traffic Monitoring System,TP277
  17. On the television program template of intellectual property protection,G222
  18. Copper oxide porous hollow microspheres Preparation and Characterization,O614.121
  19. Research on Abnormal Vehicle Behavior Detection and Application in Traffic Video,TP391.41
  20. Research on License Plate Location Algorithm of License Plate Recognition System,TP391.41
  21. Research on Face Detection Based on Skin Color Segmentation and AdaBoost Algorithm,TP391.41

CLC: > Industrial Technology > Automation technology,computer technology > Computing technology,computer technology > Computer applications > Information processing (information processing) > Pattern Recognition and devices > Image recognition device
© 2012 www.DissertationTopic.Net  Mobile