|
|
|
|
|
|
Title
|
Complex Classification Using Advanced Machine...
 |
|
|
|
Abstract |
|
In this report, we studied different complex classification models such as Gradient Descent, Multiclass Classification. We also used different Machine Learning tools such as LibSVM, MEGAm, and FastDT to design a complex classifier based on OVA and AVA approaches. We also designed 2 different Rank classifier using MEGAm library and evaluated its performance on the OHSUMED database. The binary classification accuracy (0-1) error using 20 different queries and 10 retrieved documents for each query was 33% for Ranking Classifier 1. The binary classification accuracy for Ranking Classifier 2 was 37%. However, the average ranking performance, as evaluated using DCG metric, was roughly 8% better for Ranking Classifier 2 as compared with Ranking Classifier 1. This improvement comes from the cost function used to penalize the mis-ranking. |
| |
|
| |
|
|
|
|
|
Title
|
|
A Machine Learning approach to localization ...
|
|
|
|
Abstract |
|
Many sensor network related applications require precise knowledge of the location of constituent nodes. In these applications,it is desirable for the wireless nodes to be able to autonomously determine their locations before they start sensing and transmitting data. Most existing localization algorithms rely on anchor nodes whose locations are known to determine the positions of the remaining nodes using methods such as triangulation or trilateration. In this work, we consider the scenario where anchor nodes are not equipped with any self-positioning functionality, and do not have the knowledge about their positions. In such cases, anchor nodes respond to a "HELLO" signal transmitted by the localizing node. The response to "HELLO" signal can be used by the localizing node to estimate the time of arrival (ToA) or Received Signal Strength Indicator (RSSI). We use such ToA or RSSI measurements to learn node locations using Support Vector Machines (SVM). We cast the problem into a regression and a multiclass classification setting, and demonstrate the high localization accuracy achieve by this approach as compared with the traditional Least Squares based solution. We also demonstrate that strategically choosing the evaluation order o... |
| |
|
| |
|
|
|
|
|
Title
|
|
Gender Classification Using Support Vector Ma...
|
|
|
|
Abstract |
|
Automatic face recognition is a classic problem in the area of computer vision research. This problem is still a very active area of research in vision community. The primary reason for this problem to get so much attention is the fact that face recognition finds application in many commercial applications and can work as a biometric in many law enforcement applications. The problem of Automatic face recognition can be formally defined as follows: Given a set of representative training images for each person in the database, determine the identity of a new face images from the stored data. There have been several techniques proposed in literature to extract different type of features related to shape, color, etc. of the face. Some of the techniques simply use the image pixel values as the features and reduce the dimension of these features by applying some constraints such that classification property of training image is preserved. In Project 1 of this course, we examined different kind of feature reduction method and used nearest neighbor approach as the classifier. The fundamental problem with the use of nearest neighbored approach as classifier is that the probability of error is quite high (twice as compared with Bayes classi... |
| |
|
| |
|
|
|
|
|
Title
|
|
A Comparative Study of Different Face Recogni...
|
|
|
|
Abstract |
|
Automatic face recognition is a classic problem in the area of computer vision research. This problem is still a very active area of research in vision community. The primary reason for this problem to get so much attention is the fact that face recognition finds application in many commercial applications and can work as a biometric in many law enforcement applications [1]. The problem of Automatic face recognition can be formally defined as follows: Given a set of representative training images for each person in the database, determine the identity of a new face images from the stored data. There have been several techniques proposed in literature to extract different type of features related to shape, color, etc. of the face. Some of the techniques simply use the image pixel values as the features and reduce the dimension of these features by applying some constraints such that classifcation property of training image is preserved.In this report, we discuss some of these methods which are widely used in literature and promise to exhibits good recognition accuracy. We will use Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA), Sparse representation and Random projection for the task of face recognition and ... |
| |
|
| |
|
|
|
|
|
Title
|
|
Evaluation of Phase & Magnitude based Feature...
|
|
|
|
Abstract |
|
Given a speech signal, there are two most important information that can be extracted from it. One being the linguistic information (about what is being said) and other being the speaker specific information (about who is speaking). This report is about the task of speaker recognition where the goal is to determine the speaker identity, from a group of known speaker, which closely matches with input sample. This problem become even more tough when there is limited amount of test and train data, a mismatch between the surrounding conditions while recording the test and train data, or in noisy environments. In this thesis, we consider the problem of speaker identification in noisy and bandlimited telephonic environments using the Gaussian Mixture Model approach combined with sub-band based feature extraction. We implement a sub-band based Posteriori Union Model described by Reynolds [1]. Then, we extend sub-band based approach to combine the phase based feature ModGDF and Magnitude based feature MFCC using several feature recombination techniques described in this thesis. These sub-band based feature recombination methods gives 46% identification accuracy, in best case, on NTIMIT database with little or no increase in computation. |
| |
|
| |
|
|
|
|
|
Title
|
|
JPEG Progressive mode vs Sequential mode
|
|
|
|
Abstract |
|
JPEG standard generally use two type of scan to transmit the image using entropy coding. One is known as progressive mode, while other is known as sequential mode. In sequential mode, every block of image is encoded in a zigzag fashion from low frequencies to high frequencies and transmitted. So, at the receive, we get information in a sequential manner from left to right blocks and top most block to bottom most. While in progressive mode, the encoding for all blocks are done at the same time for same frequency components. So, the DC component of each block is encoded together, followed by first frequency components of all the blocks and so on. Since, progressive mode uses the information from all blocks at the same time, it is useful to implement JPEG using progressive mode for large image sizes. As in that case, information about DC and low frequency components can give average information about image. User can then download the complete image, if he feels the need to store that image. Progressive scan is also efficient in coding because different Huffman table can be used to encode different frequency components. Hence, performance of progressive scan is expected to be better than sequential mode. |
| |
|
| |
|
|
|
|
|
Title
|
|
Digital Image Processing
|
|
|
|
Abstract |
|
RGB Channel decomposes the original image into three channels, each corresponding to Red, Green & Blue Channel. The brightest part in the decomposed images represents the highest intensity value corresponding to the respective color channel in the original image. This is well represented in the Figure 1. The Red Roses are shown very bright in R channel image while their intensity is very low in G and B channel image. Similarly, Green leaves and grass is very bright in G channel image and Blue pollens are very bright in B channel image. The Pinkish rose (One between Green rose and Red rose) has intensity distributed across each of R,G,B channel. Hence, we can conclude that any other color can be formed by mixing the three channels in appropriate proportion. So, an image can be satisfactorily represented by these three components for analysis. |
| |
|
| |
|
|
|
|
|
Title
|
|
Software issues encountered while modeling co...
|
|
|
|
Abstract |
|
This is another excerpt from my master's thesis providing an objective discourse on the issues encountered while doing CFD modeling of continuous phenomena at the micro or nano level. |
| |
|
| |
|
|
|
Go to document page