assessment and analysis on color image classification

International Journal of Advanced Research in Electronics and Communication Engineering (IJARECE)

Volume 3, Issue 11, November 2014

1440 ISSN: 2278 – 909X All Rights Reserved © 2014 IJARECE

Assessment and Analysis on Color Image

Classification Techniques of Dermatological

Ulcers

C.Dhaniya Briskillal.,B.E.,(M.E), K.Madhan Kumar.,M.E.,(Ph.D)

Abstract -- With the implementation of color image processing

methods, the image of dermatological ulcers are analyzed in

order to detect the affected area of the skin. The detection of

classification rate focus on the application of feature extraction

method that segment, classify and analyze the tissue

composition of skin lesions or ulcers. Indexing of skin ulcer

images was performed based on the statistical texture features

derived from the RGB color components. This literature

assesses the high level methology for dermatological ulcer

image classifier. The classifiers analyzed here are used for

labeling the images by the dermatologist used in training and

testing of the classifier. The classification performance rate,

coverage area of the affected skin is analyzed based on the

choice of different algorithms. Classifier uses the algorithm to

perform attribute or feature selection that generates the

candidate subsets of attributes and evaluate them by using the

training and testing schemes, thus creating the computed

values of corrected classified image rate up to 90% and

assessed coverage area of affected skin up to 0.82.

IndexTerms: Dermatological ulcers, Classification rate, Feature

extraction, Classifiers. Texture features

I INTRODUCTION

Ulcers are usually caused by deficit in blood circulation and

can be associated with arterial or venous insufficiency,

diabetics, vascular diseases, tumors, infection and certain

specific skin conditions explained in [17]. With the

widespread use of digital cameras, freehand wound imaging

for skin ulcers has become common practice in a clinical

settings but still a demand arises for practical tool for

accurate wound healing assessment describes in [3]. The

healing process of an ulcer can be divided into three phases

such as Inflammation, tissue formation and remodeling

formulated in [10].

C.Dhaniya Briskillal, Department of Communication System,

PET Engineering College, Tirunelveli, India,

K.Madhan Kumar, Department of Electronics and

Communication Engineering, Tirunelveli, India,

During the healing process the skin can change to a

transient phase of necrosis. The appearance of a wound,

lesion or ulcer provides importantclues that can help with

the diagnosis, determination of severity and the prognosis of

healing provided in [2].

The major objective of this study is to evaluate

several classification techniques of a feature or attribute

extraction and selection to determine which color and

texture features.The tissue composition of each lesions are

classified independently by an expert dermatologist based

on the color composition by different types given below :

Granulation(Red)

Fibrin(Yellow)

Necrotic (Black)

Hyperkeratosis Or Callous(White) &

Mixed Tissue Composition

In the feature extraction and selection method, set

of features extracted with large amount of data sets. In the

process of classification among the different types of skin

ulcers, the extracted features should be evaluated to

determine which are more relevant to the solution of the

problem given in [13]. Suitable algorithm is chosen for the

classifier to perform feature selection methods that generates

candidate subsets of attributes and this process is repeated

with each candidate set until the stopping criterion is

reaches.

Cross validation is used to estimate the accuracy of

the given set of attributes that are module in [5]. Cross

validation is a sampling method that randomly divides the

samples into r mutually exclusive partition or folds of

approximately equal in size of n/r samples where n is the

total number of available number of samples. The classifier

is trained with the r-1 induced folds of samples and the

classification rule or methodology is tested on the remaining

fold. This process is repeated r times, so that each fold is

used once as the test test given in the system.




The general system design describes the above processing unit is module in a block diagram given below in Fig.1

Fig 1. General system design for color image classification

The section I describe the general system of the color image classification system with the basics of different types of ulcers

with the suitable preprocessing and feature extraction method. The section II defines the color image classifiers that are used

to classify the segmented image and reproduce the classification rates. The section III defines the various types of

methodology used in the classification methods that improves the performance of classification rate. The section IV explains

the overall performance metrics of the classification rate. Finally section V concludes with the best performance of

classification method.

III COLOR IMAGE CLASSIFIERS

Different classifiers chosen to run the algorithm are

analyzed and describes based on the Fig.2

Fig.2 Types of Color Image Classifiers

In [9], the KNN classifier finds the k-nearest neighbors of

the samples to be classified by using a distance metric

(usually the Euclidean distance) between the attributes of

the sample to be classified and the attributes of all the

available samples with known classification.

In [7], Wound Image Analysis classifier is used to classify

the wounds based on the color description of various

colors involves in analyzing the severity of the wound.

WIAC framework is used efficiently to forecast the

healing process of a wound effectively which uses the

non-contact method of analyzing the status of the wound.

In [6], the Naive Bayes or Bayesian classifier calculates

the probability of a sampling belonging to each of the

predetermined classes. This probability value is used to

generate a model with a decision rule that provides a

response to indicate the class with a higher probability.

In [1], the ANN classifier is an interconnected group of

nodes that is capable of machine learning as well as

pattern recognition. It consist of set of adaptive weights

i.e.) numerical parameters that are tuned by learning

algorithm.

In [17], MLP classifier consists of a set of nodes that

constitute the input layer, one or more intermediate or

hidden layers of computational nodes and output layer.

The set of measurements to be classified is provided to the

input layer and propagates forward through the hidden

layer towards the output layer creating the computed

value for classification.




II METHODOLOGY

This section describes the analysis of different image

classification methods with suitable algorithms to predict the

best extent classification rate of a classifier discussed below:

A. ROUND-ROBIN CLASSIFICATION

This method describes the color image

classification and diagnosis of four major groups of prostate

cancer such as stroma, benign prostatic hyperplasia,

prostatic intra-epithelial neoplasia and prostatic carcinoma.

To overcome the high dimensionality of multiclass system, a

novel round robin tabu search algorithm is used to segment

image sample of four major class of prostate cancer with

classification performance of about 91.9%. Since the

classifier uses different features for each binary classifier,

the results increases for classification accuracy but the time

elapsed for the execution process. Table.1 describes the

comparison of the multiclass method and Round robin

classification of the system that results lower computational

time with the round robin classifier than the multiclass

imagery method.

Table.1 Computation Time Comparison

In round-robin algorithm, simple voting scheme is

used to predict the highest priority class among the four

major classes of prostate cancer. The samples are correctly

classified by the nearest neighbor classifier because different

feature are used for each binary classifier and results in the

increase of classification accuracy. Once the Tabu search

find the best subset of features, an Nearest Neighbor

classifier are used to determine the class of new subset but

the cost of measuring feature is increased because more

features are required in Round Robin approach. Neighbors

are calculated using a squared Euclidean distance defined as

𝑑 𝑥, 𝑦 = (𝑛𝑖=1 𝑥I – yi)

2

where x and y are two input vectors and n is the number of

features.

From the table it is inferred that the execution time

using RR classifier such as 74.49msec for Dataset1and

109.48msec for Dataset2 are higher than the Multicast

classifier such as 48.31msec for Dataset1 and 65.361 for

Dataset2 for classification process.

B.MLP BASED WRAPPER ALGORITHM

This methodology describes the color images of the

skin ulcers that are detected and classified using the MLP

classifier. First the database images are collected and each

images are independently and manually segmented into two

regions, describes the lesion and the background by an each

region attributes to RGB color components and convert it to

the HSI component. The wrapper algorithm is used to

specify the type of ulcers where this algorithm generates the

candidate subsets of attributes and evaluates them by using

HSI images. This process is repeated with each candidate set

until the stopping criterion is reached.

Fig.3 Analysis of texture in color imaging system.

Database images collected are given to the system.

In preprocessing, each images are independently and

manually segmented into two regions represents the lesion

and background. The noise is reduced by feature

extraction method. Wrapper algorithm uses the cross

validation that is used to estimate the accuracy of the

learning scheme for a given set of attributes. The wrapper

algorithm uses the MLP classifier that consists of a set of

nodes that constitute the input layer, one or more hidden

layer or output layer. The set of measurements to be

classified is provided to the input layer and propagates

forward through the output layer created a computed value

for the classification up to 73.8%.

C.COLOR PIXEL CLASSIFICATION

The color pixel classification approach describes

the color representation prediction, color quantization and

classification algorithm. Skin segmentation based on color

pixel classification is not affected by the color space but the

segmentation performance is degraded by the chrominance

channels used in the classification. Increase in histogram

size will increase the performance of segmentation. To

overcome all the difficulties, Bayesian classifier with the

Histogram technique is used. Color pixel classification using




Bayesian classifier is used to determine and differentiate the

skin color and non-skin color. The Bayesian classifier with

the histogram technique provides higher classification rates

compared to other classifiers. The Bayesian classifier with

the histogram technique is feasible for the skin color pixel

classification problem because the feature vector has a low

dimension.

Fig. 4. Sample images from the ECU face and skin detection

database. (a) Original image (Set 1). (b) Face segmented image

(Set 2). (c) Skin segmented image

(Set 3).

The Bayesian decision rule for minimum cost is a

well-established technique in statistical pattern

classification. Using this decision rule, a color pixel x is

considered as a skin pixel if

𝜌(𝑋|𝑠𝑘𝑖𝑛)

𝜌(𝑋|𝑛𝑜𝑛𝑠𝑘𝑖𝑛)≥ 𝜏

Where 𝜌(𝑋|𝑠𝑘𝑖𝑛)and 𝜌(𝑋|𝑛𝑜𝑛𝑠𝑘𝑖𝑛)are the respective

class-conditional pdfs of skin and nonskin colors and 𝜏 is a

threshold. The theoretical value of 𝜏 that minimizes the

total classification cost depends on the a priori probabilities

of skin and nonskin and various classification costs;

however, in practice 𝜏 is often determined empirically. The

class-conditional pdfs can be estimated using histogram or

parametric density estimation techniques.

Fig. 5. ROC curves of skin color pixel classifiers

The Bayesian classifier with this histogram

technique used for the skin color pixel segmentation will

provide a classification rate of 89.84% and the correct

detection rate ranges up to 91%. Although this technique

provides good coverage of all different skin types, it lacks of

available pdf of datasets. Skin color pixel luminance is

discarded due to the different lighting variation which made

more complexity in computational time. Increase in

histogram size leads to the finer Pdf estimation to attain the

high accuracy in classification rate.

D. CLUSTERING TECHNIQUE

This technique describes the k-means clustering

algorithm followed by the Artificial Neural Network

classifier that is used for the skin ulcer detection at an early

stage. Since skin ulcers are caused due to the increase in the

UV radiation, if detected earlier all forms of skin cancers are

curable. Thus this paper depends on the basis of classifying

skin lesion at an early stage.

Fig.6. Objective function comparison with the number of

iterations of extracted features.

Due to the complexity and irregular variations of

skin visual inspection of skin lesion may not be properly

classified and detected at an early stage. To avoid this,

proper classification algorithm should be chosen. This paper

proposed the k-means clustering technique that select best

peak of an image and set limits on each of the dimensions

for each separate region. This makes it easier to detect the

skin ulcer and heal the patient rather in early stage. By using

this algorithm, will provide ability to handle high

dimensional data and does not requires any domain

knowledge. ANN classifier used here will provide a

classification rate of skin lesion up to 85 to 90%, thereby

helps the dermatologist to detect and diagnose the problems.

Provides more accuracy of classification rate than other

classifier method but insists more computational time due to

the classification of relatively similar colors that makes the

feature extraction more difficult.

E.WIAC ALGORITHM

This WIAC algorithm describes the status of the

wound healing assessment that measures the area covered

by the wound. WIAC is an efficient classifiers that are used

to classify the wound images. The purpose of this paper is to

accurately access the healing status of the wound that




captures the wounded images by the use of tools like

photographic wound assessment tool to predict the different

types of wounds such as pressure, diabetic and arterial

ulcers. After acquiring the required image of wound, it leads

to the segmentation process that segments the wounded area

from that of unwounded area and pre-processing is done to

reduce the noise using efficient filtering and denoising

techniques.

Fig.7. Flowchart of Wound Image Analysis Classifier

WIAC Algorithm Steps:

Select a wound image from wound database acquired

from open source wound images.

Segment the wound from the wound image using an

efficient segmentation technique.

Improve the quality of the segmented wound from the

wound image by using filtering,

Denoising techniques.

After preprocessing of segmented wound image

overlay transparent layers of segmented

Wound shapes to reduce the intensity of colors in the

wound.

Repeat step4 various times to get healed image of the

wound.

Classify the segmented wound image into three labels

based on severity level 0,1,2.

Conversion of segmented wound images from higher

severity level to lower severity take place.

Level leads to analyzing the healing status of the

wound.

WAIC algorithm provides the conversion of

segmented wound images from higher severity level to

lower severity level that leads to analyze the healing status

of wound images. The classification time during execution

ranges higher than other classifier. The percentage of

classification accuracy range with a value of 63.8% due to

the usage of photographic wound assessment tool used in

detection of grade, number and location of wound that will

not be indexed properly

F.LOG POLAR FOURIER TRANSFORM

An integrated computer based system for the

characterization of skin digital images was described by the

image acquisition arrangement designed for capturing skin

images, under reproducible conditions. This system

processes the captured images and performs unsupervised

image segmentation and image registration utilizing an

efficient algorithm based on the log-polar Fourier transform

of the images by means of the ANN classifier. Cross talking

between neighboring pixels is a factor that contributes to

noise. Noise can be eliminated through appropriate filtering

such as morphological filtering to preserve the edge

information or median filtering to assist the color

identification. The human skin consists of irregular surface,

scatters incident light in irregular directions and burying

diagnostic information. To avoid such problems the

appropriate lighting geometry, light reflections reduction

and software corrections also applied to the image during

the acquisition.

Fig.8. Generic feed-forward neural network used for

classification.

If the two images are𝑓(𝑥, 𝑦) and𝑓 ′(𝑥, 𝑦) , they are

connected through a four-parameter geometric

transformation, which is described by the following:

𝑓 ′ 𝑥, 𝑦 = 𝑓(𝑎 𝑥𝑐𝑜𝑠𝑏 + 𝑦𝑠𝑖𝑛𝑏 ∆𝑥, 𝑎 −𝑥𝑠𝑖𝑛𝑏 + 𝑦𝑐𝑜𝑠𝑏 − ∆𝑦)

where ∆𝑥 and∆𝑦 are the parallel shifting,𝑎 is the

magnification factor, and 𝑏 is the rotation angle.

The Fourier spectra of the two images are

connected by the following:

𝐹′ 𝑢, 𝑣 =1

𝑎2|𝐹

(𝑢𝑐𝑜𝑠𝑏 + 𝑣𝑠𝑖𝑛𝑏)

𝑎 ,

(−𝑢𝑠𝑖𝑛𝑏 + 𝑣𝑐𝑜𝑠𝑏)

𝑎 |

which displays the independence of horizontal and vertical

shifting. Then the log-polar transformation is performed

𝐹′ 𝑟,𝜃 =1

𝑎2|𝐹

𝑟

𝑎 ,𝜃 + 𝑏 |




Where,

𝑟 = 𝑢2 + 𝑣2

,𝜃 = 𝑡𝑎𝑛−1𝑣

𝑢

To create the uniform illumination at an angle

greater than 45° over the entire field of view was difficult to

deficted. Image registration method used for monitoring the

progess and measuring the changes of skin lesions. The

major advantage of this image acquisition handling system

is their easy implementation. The Fourier transformation

spectrum with log polar transforms is independent of

horizontal and vertical shifting used to eliminate the

dependency on magnification and rotation. The log polar

transform had good performance in measuring the small

rotation up to 10°. For larger rotations, the algorithm

calculates the greater rotation values than the real ones.

Fig.7 shows the generic feed forward neural network used

for the classification by the log polar transform is predicted

by the above equations. The classification rate predicted by

ANN classifier ranged up to 85%. The ANN classifier used

maps a large number of inputs into a small number of output

to detect the characteristics measure of pigmented skin

lesions, therefore the computation time increases.

G.BACK PROPOGATION ALGORITHM

This paper describes the detection of skin burn and

classification of images using Artificial Neural Network

classifier. The images are collected from various database

centre and it is preprocessed by enhancing the lab color

space and then various pattern analysis such as ANN

classifier is applied on the collected skin burn images to

predict the classification rate for three types of grade such as

Grade 1 represents the superficial burn caused due to skin

burn heals within 5 to 7 days. Grade 2 represents the partial

thickness burn that leaves scars on the affected area. Grade 3

represents the full thickness burns that are totally destroying

the epidermis layer of the skin. These proposed systems are

more useful at the remote location where the medical

experts are not available. During the accidents as the

patients gets admitted to the hospital or while being

transported to ambulance, the medical personnel attending

the patient can send burn images immediately through

online camera to the specialist.

Feature selection is very important while

classifying the skin burn image into different grades.

Initially the image is re-sized to 90*90 pixels, and then Red,

Green and Blue (RGB) space is converted in to L*a*b*

color space. Lab color space has 3 coordinates, one

luminous and two chrominance V1 and V2. Luminous

component is used for contrast enhancement. After contrast

enhancement of the image, the V1 and V2 chrominance

planes of the L*a*b* color space is selected for feature

extraction using the coefficient of Discrete Cosine

Transform (DCT) function is chosen to train classifiers. The

two dimensional DCT equation is given below,

where 𝑋(𝑘1 , 𝑘2) is the DCT an 𝑥(𝑛1,𝑛2) is the image

𝑋(𝑘1 , 𝑘2) =4 ∈𝑘1

∈𝑘2

𝑁2 𝑥(𝑛1,

𝑁2−1

𝑛2

𝑁1−1

𝑛1

𝑛2)

cos(𝜋 2𝑛1 + 1 𝑘1

2𝑁1

)cos(𝜋 2𝑛2 + 1 𝑘2

2𝑁2

)

Where k=1, 2, 3……….. N

Fig.9. Flow Chart for Training ANN With Burn Images

Using BPA

The ANN classifier is used with the Back

Propagation Algorithm to segment the severity of the skin

burn wounds. Fig.8 describes the Back Propagation

Algorithm that convert the collected database into a required

format to extract the features from the color information.

Finally by using Artificial Neural Network classifier, it will

display the grades of burns and provide the reduction of

classification time without any complex algorithm. It

predicts the classification rate of order of 90% for Grade-1,

75% for Grade-2 and 82.5% for Grade-3. It provides high

classification rate for Grade-1 when compared to Grade-2

and Grade-3 since Grade-1 refers to simplify the top most

layer of the skin.

III OVERALL PERFORMANCE

This section describes the overall performance of the

classification rate of different classifier analyzed in the

above sections. Fig.9 shows the classification rate of

different classifier.The MLP classifier described in[17] will

provide a classification rate of 73.8% whereas KNN

classifier described in [9] uses different features for each

binary classifiers, the results increases for classification

accuracy with the computed value of 91.9% .




The ANN classifier assessed in [1] describes the

skin ulcer detection at an early stage with the use of the k-

means clustering algorithm The Bayesian classifier analyzed

in [6] describes the color pixel classification used with the

histogram technique provides the classification rate of

89.84% but it increase the computational time. By using the

histogram technique, color space favours over other colors,

that overlap between skin and nonskin types. Increase in

histogram size leads to finer pdf estimation but requires

greater memory storage.

Fig.10.Classification performance analysis of different

classifier

The WIAC classifier described in [7] used to

predict the status of wound healing assessment.The

percentage of classification accuracy range with a value of

63.8%. From the above analysis graph in Fig.10, it is infer

that KNN provide the best classification rate among all the

classifier but the computational complexity increases will

further affects the classification accuracy. The MLP

classifier used with the wrapper algorithm provide the

reduced computational time which will improve the

classification accuracy to the better extend.

IV CONCLUSION

The generation and estimation of tissue composition

provides key information for monitoring the response to

treatment of patients with dermatological ulcers. This

processing system may serve as a diagnostic adjacent for

medical professionals for the confirmation of a diagnosis of

dermatological ulcer as well as for the training of new

dermatologists. Various classifiers analyzed in this paper

will provide the implementation of more accurate, faster and

more reliable classification system. Six different algorithms

are used for image segmentation along with different

classifiers provides the better extent of classification rate.

Out of the five classifiers tested, KNN classifier provided

the best overall performance with the classifier rate value

ranges of 91.9%.

The results from this survey are promising for the

future that an objective analysis of color image classification

of skin ulcers using proposed method with the SVM

classifier can overcome some of the limitation of visual

analysis, storage requirements and lead to the development

of improved optimization for the treatment and monitoring

of skin ulcers with the high percentages of corrected

classification rate.

REFERENCE

[1] A. Tarallo, A. Gonzaga, and M. Frade, “Artificial Neural

Networks Applied to the Segmentation and

Classification of Digital Images of Cutaneous Ulcers,”

In Proc. IEEE 7th Int. Conf. Bioinformatics. Bioeng.,

2007, Pp. 1–1.

[2] Goldman, R.J., Salcido, R”Processing Of Digital Images

Of Cutaneous Ulcers Classification Using MLP”.

International Journal of Dermatology,

Vol.44,Issue.3,November 2012.

[3] Hazem Wannous, Yves Lucas*, Member, IEEE, And

Sylvie Treuillet” Enhanced Assessment Of The Wound-

Healing Process By Accurate Multiview Tissue

Classification”, IEEE Transactions On Medical

Imaging, Vol. 30, No. 2, February 2011.

[4] H. Oduncu, A. Hoppe, M. Clark, R. Williams, And K.

Harding, “Analysis Of Skin Wound Images Using

Digital Color Image Processing: A Preliminary

Communication,” Int. J. Low Extreme Wounds, Vol. 3,

No. 3, Pp. 151–156, 2004.

[5] H.Wannous, S. Treuillet, Andy. Lucas, “Robust Tissue

Classification for Reproducible Wound Assessment in

Telemedicine Environments,” J. Electron. Imag., Vol.

19, P. 023002, 2010.

[6] I. Maglogiannis, Member, IEEE, S. Pavlopoulos,

Member, IEEE, And D. Koutsouris, Senior Member,

IEEE “An Integrated Computer Supported Acquisition,

Handling, and Characterization System for Pigmented

Skin Lesions In Dermatological Images”, IEEE

Transactions On Information Technology In

Biomedicine, Vol. 9, No. 1, March 2005

[7] K. Sundeep Kumar And B. Eswara Reddy“Wound Image

Analysis Classifier For Efficient Tracking Of Wound

Healing Status” Signal & Image Processing : An

International Journal (SIPIJ) Vol.5, No.2, April 2014

DOI : 10.5121/Sipij.2014.5202 15.

[8] Muhammad Atif Tahir, Member, IEEE, And Ahmed

Bouridane, Senior Member, IEEE “Novel Round-Robin

Tabu Search Algorithm For Prostate Cancer

Classification And Diagnosis Using Multispectral

Imagery”, IEEE Transactions On Information

Technology In Biomedicine, Vol. 10, No. 4, October

2006.

0

20

40

60

80

100

Classification analysis of

different classiifier

MLP

KNN

ANN

BAYESIAN

WIAC

LOG POLAR

BPA




[9] Malini Suvarna, Sivakumar, Dr. Kamal Kumar & U C

Niranjan “ Diagnosis Of Burn Images Using Template

Matching, K-Nearest Neighbor And Artificial Neural

Network”, International Journal Of Image Processing

(IJIP), Volume (7) : Issue (2) : 2013.

[10] M. Witte and A. Barbul, “General Principles of Wound

Healing,” Surg. Clin. North Amer., Vol. 77, No. 3, Pp.

509–528, 1997.

[11] Patel Jigna J, Bhatt Bhumika “Review On Detection

Of Skin Ulcers At An Early Stage “, International

Journal Of Emerging Technology And Advanced

Engineering Website: Www.Ijetae.Com (ISSN 2250-

2459, ISO 9001:2008 Certified Journal, Volume 4,

Issue 3, March 2014)

[12] P. Domingos And. J. Pazzani, “On the Optimality of

the Simple Bayesian Classifier under Zero-One Loss,”

Mach. Learn., Vol. 29, No. 2, Pp. 103–130, 1997.

[13] R. Kohavi, “A Study of Cross-Validation and Bootstrap

for Accuracy Estimation and Model Selection,” In Proc.

Int. Joint Conf. Artif. Intell., 1995, Vol. 14, Pp. 1137–

1145.

[14] //http://en.wikipedia.org/wiki/Support_vector_machine

[15] S. M. Pereira, M. A. C. Frade, R. M. Rangayyan, and P.

M. De Azevedo Marques. (2011). Classification of

Dermatological Ulcers Based On Tissue Composition

and Color Texture Features. In Proc. 4th Int. Symp.

Appl. Sci. Biomed. Commun. Technol., Ser. ISABEL

2011

[16] Son Lam Phung, A. Bouzerdoum , D. Chai “Skin

Segmentation Using Color Pixel Classification:

Analysis And Comparison”, IEEE Transactions On

Pattern Analysis And Machine Intelligence, January

2005, 27(1), 148-154.

[17] Silvio M. Pereira, Marco A. C. Frade, Rangaraj M.

Rangayyan, Fellow, IEEE, and Paulo M. Azevedo-

Marques, Member, IEEE” Classification Of Color

Images Of Dermatological Ulcers”, IEEE Journal Of

Biomedical And Health Informatics, Vol. 17, No. 1,

January 2013.

assessment and analysis on color image classification

Documents