16-09-2014, 03:16 PM
Salient point detection
Salient point detection.docx (Size: 452.04 KB / Downloads: 9)
INTRODUCTION:
Object recognition is of greater task in computer vision. It is the concept of finding objects in an image or video sequence. The object recognition has extended their applications in many areas such as Image Panoramas, Image Watermarking, Global Robot Localization, Face Detection, Optical Character Recognition, Content-Based Image Indexing Automated Vehicle Parking Systems and Visual Positioning and Tracking[1]. Normally, Human visual systems recognize large number of objects from images with little effort even when they have different appearances in different circumstances. But computer systems face it with lot of difficulties to recognize objects that are occluded by other objects and that having different appearances [2]. The proposed method overcomes this strategy that it recognizes objects in these challenging circumstances and provides better results on Caltech database.
1.1 Salient point detection:
Salient point detection plays an important role in content based image retrieval in order to represent the local properties of the image. Since classic comer detectors cannot support natural images, detector based on wavelet transform to represent global as well as local features is used to detect the salient points [3, 4]. C.Schmid et al used local gray invariants for image retrieval where local gray invariants are computed automatically over the detected interest points [5]. Weber et al proposed the computation of K-means clustering algorithm at Forstner interest points for object recognition [6].
Patches have to be extracted at different scales to address the scale difference of the objects. Moreover occlusions of images can easily be handled by these patches [7]. Alexandra Teynor et al calculated gray values, Haar integral gray invariants and SIFT for image patches over the interest points for visual object class recognition [8].
Due to mathematical and biological properties, Gabor Wavelet Features are applied in Object Representation, Face Recognition, Tracking and Edge Detection [9]. The views of
Gabor Features for various applications vary from author to author.
The work based on Gabor wavelet for face recognition is proposed by Shiguang Shan et al [10]. The work is done by implementing face recognition system based on discriminate
Analysis of Gabor representation. P.Kruizinga et al compared and evaluated the texture features using local spectrum obtained by a bank of Gabor filters [11]. Xiao gang Wang et al
Combined Bayesian Probabilistic model with Gabor features for face recognition [12]. where feature vectors are extracted based on dominant orientation matrix computed from the bank of Gabor filters [13]. The work done on object recognition using Gabor wavelet is proposed by SHEN Lin-Lin et al [14]. The work is based on extracting features using Gabor wavelet with 5 scales and 8 orientations and classifying features using SVM classifier for face recognition.
The main goal is to improve the literature survey by using effective features in order to handle occluded part of the image. The contribution of this work is,
(i) Salient Point Detection and Patch Extraction to overcome the strategy of locating the objects present in the particular part of the image.
(ii) Feature Extraction to handle various complex images which looks different in different circumstances. A fair amount of work has
been done previously for face recognition using Gabor wavelet features with different scales and orientations. In this proposed method, the whole Caltech database is evaluated based on
Gabor wavelet features with 2 scales and 2 orientations and 2 scales and 4 orientations. Gabor wavelet features highly discriminate the images irrespective of their localizations in
different circumstances. Then these features are classified using SVM classifier. The single class object recognition task that is to recognize object or non object has been done here to
Increase the robustness of the proposed method. Thus the task evaluated using performance measures and error rate on different categories of Caltech database support proposed method with better results.
1.2 Feature extraction:
Feature extraction and classifier learning are essential to the performance of an object recognition system. Discriminative features and robust classifiers are always desirable to pattern recognition applications. However, discriminative features like Gabor features, either require computationally expensive feature extraction process, or have large feature dimension [1]. The large feature dimension could bring huge computation and memory cost to the following classifier training and classification, thus makes robust classifiers, e.g Support Vector Machine (SVM), inapplicable. This paper tries to propose a framework to design efficient and robust object recognition system. Due to its biological similarity to human vision system, Gabor features have been widely used in object recognition applications like fingerprint recognition [2] and character recognition [3] etc. One of the most important object recognition applications, face recognition, has also seen the advantages of Gabor features based systems over many other methods in the literature. For example, the Elastic Bunch Graph Matching (EBGM) algorithm [4] has shown very competitive performance and was ranked the top performer in the FERET evaluation [5]. In a recent face verification competition (FVC2004), both of the top two methods used Gabor wavelets for feature extraction. Chung et al. [6] use the Gabor features over a set of 12 fiducially points as input to Principal Component Analysis (PCA) algorithm, yielding a feature vector of 480 components. They claim to have improved the recognition rate up to 19% with this method compared to that by a raw PCA. Liu et al. [7] factorize the Gabor responses and then apply a down sampling by a factor of 64 to reduce the computation cost of the following subspace training. Their Gabor-based enhanced Fisher linear discriminate model outperforms Gabor PCA and Gabor fisher faces. A more detailed survey on Gabor wavelet based face recognition methods can be found in [1].Despite the success of Gabor feature based object recognition systems, both the feature extraction process and the huge dimension of Gabor features extracted demands large computation and memory costs, which makes them impractical for real applications [1]. For the same reason, SVM has seldom adopted Gabor wavelets for feature extraction. While subspace methods like PCA and LDA could be applied for dimension reduction [7, 8], they do not improve the efficiency of feature extraction process. Some works in the literature have tried to tackle this problem by (1) down sampling the images [9], (2) considering the Gabor responses over a reduced number of points [6], or (3) down sampling the convolution results [7, 8]. Strategies (2) and (3) have also been applied together [10]. However, down sampling methods suffer from a loss of information because of the down sampling, or dimension reduction. Furthermore, the feature dimension after down sampling might still be too large for the fast training of SVM. To make SVM applicable to Gabor features, Qin and He [10] reduced the size by including only the convolution results over 87 manually marked landmarks. However, locating the 87 landmarks itself is a difficult problem. Furthermore, our works [11] have also shown that facial landmarks like eyes, nose and mouth might not be the optimal locations to extract Gabor features for face recognition. In this paper, we propose a general object recognition framework based on SVM and optimized Gabor features. The most significant positions for extracting features for face recognition are first learned using a boosting algorithm, where the optimized Gabor responses are computed and used to train a two-class based SVM for classification. Since only the most important features are used, the two-class SVM.
RESULTS AND DISCUSSIONS:
There are 1074 airplane images, 1155 car images, 826 bike images, 450 face images and 186 leaf images in the Caltech database. Along with this there are 900 mixed background
Images commonly used as negative images for all four categories except car and separate 1370 background images containing only roads and street scenes which are used as negative images for car category. The images in Caltech database are of various sizes and for the experimental purpose; they are converted into gray scale images [17].
Initially 250 most prominent points are taken using wavelet based salient point detector for every image. The patch of size 11 x 11 is extracted over each salient point. Then Gabor wavelet features such as Gabor mean and variance are extracted for each and every patch that extracted from the original images over the salient points. Finally, the features
thus obtained are given to SVM classifier in order to recognize objects based on these Gabor wavelet Features.