SELF-TAUGHT LOW-RANK CODING

FOR VISUAL LEARNING

ABSTRACT:
The lack of labeled data presents a common challenge in many computer
vision and machine learning tasks. Semisupervised learning and transfer
learning methods have been developed to tackle this challenge by
utilizing auxiliary samples from the same domain or from a different
domain, respectively. Self-taught learning, which is a special type of
transfer learning, has fewer restrictions on the choice of auxiliary data. It
has shown promising performance in visual learning. However, existing
selftaught learning methods usually ignore the structure information in
data. In this paper, we focus on building a self-taught coding framework,
which can effectively utilize the rich low-level pattern information
abstracted from the auxiliary domain, in order to characterize the high-
level structural information in the target domain. By leveraging a high
quality dictionary learned across auxiliary and target domains, the
proposed approach learns expressive codings for the samples in the
target domain. Since many types of visual data have been proven to
contain subspace structures, a low-rank constraint is introduced into the
coding objective to better characterize the structure of the given target
set. The proposed representation learning framework is called self-taught
low-rank (S-Low) coding, which can be formulated as a nonconvex
rank-minimization and dictionary learning problem. We devise an
efficient majorization–minimization augmented Lagrange multiplier
algorithm to solve it. Based on the proposed S-Low coding mechanism,
both unsupervised and supervised visual learning algorithms are derived.
Extensive experiments on five benchmark data sets demonstrate the
effectiveness of our approach.

ARCHITECTURE:
EXISTING SYSTEM:

The performance of visual learning algorithms is heavily dependent on

the quality of data representation. Sparse coding dictionary learning],
and low rank learning have been widely used for representing visual
data. Good representations are expressive, meaning that a reasonably
sized dictionary (i.e., basis functions) can capture a huge number of
possible input configurations, and also characterize a given set of data
with certain global structural blueprint (i.e., multiple clusters, subspaces,
or manifolds). However, the lack of training data presents a common
challenge in many sophisticated representation learning algorithms .
Traditionally, this problem was partially addressed by semisupervised
learning or transfer learning methods. Semisupervised learning makes
use of some labeled samples and a larger set of unlabeled samples,
which are drawn from the same domain with the same distribution, to
train a model. In other words, semisupervised learning can only solve
learning problems in the same domain. In transfer learning, this
restriction is relaxed to some extent. The labeled samples and the
auxiliary samples in transfer learning are drawn from different domains
with different distributions. But transfer learning requires that two
domains should be similar to each other. Many transfer learning
algorithms assume that two domains share a similar knowledge
structure. In a word, both semisupervised learning and transfer learning
put strong restrictions on auxiliary (source) data, which limited their
applicability. Recently, an emerging machine learning paradigm of self-
taught learning (STL) using unlabeled data with fewer restrictions holds
significant promise in terms of enhancing the performance of image
clustering and classification.

DISADVANTAGE:

 With the help of rich information from the auxiliary domain, we

learn effective feature representations, S-Low codings, by
incorporating low-rank constraints in the target domain.
 The proposed STL approach is a general framework that is
suitable for various visual learning scenarios. In this paper, we
present the detailed algorithms for unsupervised learning and
supervised learning.
 Instead of using the biased estimators like norm and nuclear norm
in many low-rank representation (LRR) algorithms, we replace the
l0 norm and the rank function in our model by the minimax
concave penalty (MCP) norm and the matrix γ -norm that are
considered as unbiased estimators. An effective majorization–
minimization optimization algorithm is developed to solve our
model. We also empirically illustrate the convergence property of
the optimization algorithm.
 Extensive experimental results on five benchmark data sets show
that our approach consistently outperforms several representative
low-rank learning and STL methods

PROPOSED SYSTEM:
In many real-world visual learning tasks, the assumption of sufficient
training data may not always hold. Thus, involving additional data
resources to overcome the shortage of training data becomes an
important problem. Most representative solutions include
semisupervised learning and transfer learning. The former solution
addresses this problem by using a large amount of unlabeled data
from the same domain with the same distribution to build better
classifiers, while the latter one tries to leverage labeled data from
related homogenous tasks. However, neither unlabeled data with the
same distribution nor labeled data from homogenous tasks are easy to
get. Recently, there has been a surge of interest in the topic of STL,
by involving unlabeled data without the above restrictions. Raina et
al. first proposed the concept of STL by applying sparse coding
mechanism to construct a higher level representation from the
unlabeled data. Lee et al. extended Raina’s work by presenting a
generalization of sparse coding module, which could be suited to
model other data types drawn from any exponential family
distribution. From the application point of view, Dai et al. proposed a
clustering algorithm in the spirit of STL by allowing the feature
 representation from the auxiliary data to influence the target data
through a common set of features. Kuen et alemployed the core idea
of STL, and transferred stacked auto encoders for visual tracking.

ADVANTAGES:

 They only focus on a single domain, while our approach seeks help
from the auxiliary domain and
 Existing work like [30] and [42] learns a dictionary only from the
target domain, and all other existing low-rank methods do not learn
dictionaries. However, our approach learns a dictionary from both
auxiliary and target domains in the STL setting.

MODULES:

1. Evaluation using Image

The Faculty will upload the photos. Those photos can be

evaluated with the help of S-Low Clustering. It can be achieved
with the help of python’s opencv library. The matrix form is
created with answer key to identify and give the result as per the
photos.
 2. Optimization

we design a majorization–minimization augmented Lagrange

multiplier (MM-ALM) algorithm . We first introduce the generalized
shrinkage operator Dγ ,W and the generalized singular value
shrinkage operator are nonnegative matrices. The MM-ALM
algorithm contains an inner loop and an outer loop. In each iteration,
the outer loop utilizes the locally linear approximation (LLA) of the
original nonconvex problem, and forms a weighted convex problem
for optimization. In the inner loop, we adopt the inexact augmented
Lagrangian multiplier (ALM) algorithm.

3. Clustering Results
The results from the above module are handled by some math
functions to put those values into calculations. Get the total marks
accomplished by students and average of the photos can be calculated by
the auto functionalities and display to users.
4. Data Sets and Settings

randomly select 5000 unlabeled images from the LabelMe to

construct the sample set in auxiliary domain. Currently, the LabelMe
data set contains more than 100 thousand images collected from various
resources, which provide us a great auxiliary domain for STL. shows
some images in the LabelMe data set. To evaluate how the data size in
the auxiliary domain affects the performance of learning tasks in the
target domain, we alter the number of auxiliary samples from 1000 to
5000, and compare the performance in different settings. In our
experiments, we find that increasing the size of auxiliary sample set
would improve the performance of learning tasks, but the improvements
are marginal when the size is over 3000. Due to the space limit, we only
report the results of STL algorithms under two settings that use 1000 and
3000 auxiliary images, respectively

Target Domain Data Sets: To extensively testify our approach and

related methods, we utilize the following five benchmark data sets.

ALGORITHM
When label information is available in the target domain, we design a
classification algorithm based on our S-Low coding approach to train a
classifier. Then, with the help of the learned dictionary D, our algorithm
could classify new test samples. As discussed in Section III-A, low-rank
codings ZT can be considered as new representations of the target
sample set XT . Given a test sample y, we can calculate the
representation coefficients of y ∈ Rd×1 by solving
min a y − Da2 2 + λa1

where a ∈ Rr×1 is the coefficient vector of y over D.

Input: data matrix X = [XS XT], nearest neighbors K, number of clusters

C
1: Obtain the low-rank representation matrix ZT using Algorithm 1;

2: Build an undirected graph G based on ZT (using (19)), where the

edges are weighted using correlation coefficients of each pair of
samples;

3: Prune graph G by removing some edges with small weights (keep K

nearest neighbors for each node);

4: Use NCut to generate C clusters. Output: clustering index vector L

Input: data matrix X = [XS XT], class labels of XT, test sample y

1: Obtain the low-rank representation ZT and dictionary D using

Algorithm 1;

2: Train a SVM classifier using ZT;

3: Calculate sparse representation of y using (20);

4: Predict class label of y. Output: predicted class label cy

REQUIREMENT ANALYSIS

The project involved analyzing the design of few applications so as

to make the application more users friendly. To do so, it was really
important to keep the navigations from one screen to the other well
ordered and at the same time reducing the amount of typing the user
needs to do. In order to make the application more accessible, the
browser version had to be chosen so that it is compatible with most of
the Browsers.

REQUIREMENT SPECIFICATION

Functional Requirements

 Graphical User interface with the User.

Software Requirements

For developing the application the following are the Software

Requirements:

1. Python

2. Django

3. MySql

4. MySqlclient

5. WampServer 2.4

6. Datasets(Image)

Operating Systems supported

1. Windows 7

2. Windows XP
 3. Windows 8

Technologies and Languages used to Develop

1. Python

Debugger and Emulator

 Any Browser (Particularly Chrome)
Hardware Requirements

For developing the application the following are the Hardware

Requirements:

 Processor: Pentium IV or higher

 RAM: 256 MB
 Space on Hard Disk: minimum 512MB

CONCLUSION

proposed a novel self-taught low-rank coding approach for visual

learning. Our approach jointly learned a dictionary by virtue of rich
information from auxiliary domain, and robust LRRs for target domain.
We derived both unsupervised and supervised learning algorithms for
subspace clustering and image classification, respectively. Experimental
results on five benchmark data sets demonstrated the effectiveness of
our algorithms compared with the stateof-the-art STL methods. There
remain several interesting directions for our future work:
1) given a training set in target domain, we may automatically choose
samples from the auxiliary domain and

2) we would provide fast solutions to our framework by using the

divide-and-conquer technique.