A Modified All-and-One Classification Algorithm Combined with the

Bag-of-Features Model to Address the Food Recognition Task

Kalliopi Dalakleidi, Myriam Sarantea and Konstantina Nikita

School of Electrical and Computer Engineering, National Technical University of Athens, Iroon Polytechniou Str.,

Athens, Greece

Keywords: Diabetes, All-And-One, Bag-Of-Features, Food Recognition System.

Abstract: Dietary intake monitoring can play an important role in reducing the risk of diet related chronic diseases.

Automatic systems that support patients to count the nutrient contents, like carbohydrates (CHO), of their

meals, can provide valuable tools. In this study, a food recognition system is proposed, which consists of two

modules performing feature extraction and classification of food images, respectively. The dataset used con-

sists of 1200 food images split into six categories (bread, meat, potatoes, rice, pasta and vegetables). Speeded

Up Robust Features (SURF) along with Color and Local Binary Pattern (LBP) features are extracted from the

food images. The Bag-Of-Features (BOF) model is used in order to reduce the features space. A modified

version of the All-And-One Support Vector Machine (SVM) is proposed to perform the task of classification

and its performance is evaluated against several classifiers that follow the SVM or the K-Nearest Neighbours

(KNN) approach. The proposed classification method has achieved the highest levels of accuracy (Acc = 94.2

%) in comparison with all the other classifiers.

1 INTRODUCTION

Diet related chronic diseases, such as obesity and di-

abetes mellitus, are expanding nowadays. Therefore,

an urgent need for dietary intake monitoring arises

that can reduce the risk of these diseases. Studies have

shown that when patients with diabetes mellitus do

significant errors in reporting their dietary intake,

there is an increased risk of postprandial hypo- or hy-

perglycemia. Automatic systems, usually based on a

mobile phone, can support patients that suffer from

diet related chronic diseases with carbohydrates

(CHO) counting. The user first takes a photograph of

the upcoming meal with the camera of his mobile

phone. Then, the image is processed so that the dif-

ferent types of food are divided from each other and

segmented in different areas of the image. A series of

features are extracted from each segmented area and

are fed to a classifier, which decides what kind of

food is represented by each segmented area. Then, the

volume of each segmented area is calculated and the

total CHO of the depicted meal are estimated.

Feature extraction can play a key role in dietary

intake monitoring systems. Efficient feature de-

scriptors could ensure stability and distinctiveness,

where stability means that the extracted features are

invariant to different photometric and geometric

changes and distinctiveness means that the extracted

features can be used to distinguish the specified ob-

ject from other objects or the background. Features

related to color and texture have been shown to ensure

stability and distinctiveness. Moreover, a large vari-

ety of local feature descriptors has been proposed in

the literature, like Gaussian derivatives (Florack et

al., 1994), moment invariants (Mindru et al., 2004),

complex features (Baumberg, 2000; Schaffalitzky

and Zisserman, 2002), steerable filters (Freeman and

Adelson, 1991), and phase-based local features (Car-

neiro and Jepson, 2003). A variant of Scale Invariant

Feature Transform (SIFT), Speeded Up Robust Fea-

tures (SURF), has the ability to capture spatial inten-

sity patterns, while being robust to small defor-

mations or localization errors and is shown to outper-

form the previous mentioned categories of features

(Mikolajczyk and Schmid, 2003; Bay et al., 2008).

Classification results of food images in dietary in-

take monitoring systems can be improved when the

dimension of the extracted feature vector is reduced.

The use of the Bag-Of-Features (BOF) model (Peng

et al., 2016), which is inspired by the Bag-Of-Words

model for text classification (Cruz-Roa et al., 2011)

has been reported to highly improve classification ac-

curacy in food recognition tasks. The BOF model

achieves dimensionality reduction by creating from

284

Dalakleidi K., Sarantea M. and Nikita K.

A Modiﬁed All-and-One Classiﬁcation Algorithm Combined with the Bag-of-Features Model to Address the Food Recognition Task.

DOI: 10.5220/0006141302840290

In Proceedings of the 10th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2017), pages 284-290

ISBN: 978-989-758-213-4

the extracted features visual words, and by describing

the image content with the distribution of these visual

words (Wang et al., 2016).

Another important aspect of the food recognition

task is that it is usually a multiclass classification

problem, as the used food datasets almost always con-

tain more than two categories of food. There exist

many classification approaches in order to address the

multiclass recognition task, but the most prominent

ones, like the One-Against-All (OAA), the One-

Against-One (OAO) and the All-And-One (A&O)

(Pedrajas and Boyer, 2006) descend from the binari-

zation strategy, where the division of the initial mul-

ticlass problem to several binary class problems takes

place (Galar et al., 2011).

Several attempts to implement automatic or semi-

automatic systems for dietary intake monitoring have

been reported in the literature. A food identification

application called DietCam has been recently pre-

sented (Kong and Tan, 2012), which consists of three

parts: image manager, food classifier and volume es-

timator. Images taken by the users are fed to the im-

age manager, then SIFT features are extracted, clus-

tered into visual words and fed to a simple Bayesian

probabilistic classifier, which achieves high levels of

accuracy (92%). The food volume estimator calcu-

lates the volume of each food item recognized by the

food classifier and then the calorie content of the food

is estimated. Another food recognition application

has been recently proposed for the classification of

fast-food images (Shroff et al., 2008). After segmen-

tation of the fast-food image, color, size, texture,

shape and context-based features are computed and

fed to a feed-forward artificial neural network achiev-

ing a 90% accuracy. Moreover, a food identification

system has been presented which consists of the fol-

lowing modules: image segmentation, feature extrac-

tion, food classification, and volume estimation (Zhu

et al., 2010). Food description is based on a set of

color and texture features, while classification is

based on a Support Vector Machine (SVM) classifier,

which has achieved high classification accuracy

(95,8%). An automated Food Intake Evaluation Sys-

tem (AFIES) has been reported (Martin et al., 2009),

which consists of reference card detection, food re-

gion segmentation, food classification and food

amount estimation modules. The color RGB data are

used as feature vectors for classification, which is per-

formed using the Mahalanobis distance of pixels from

food classes. The amount of calorie intake is esti-

mated based on the assumption that food area is line-

arly proportional to the food volume. In another

study, recognition of seven broad categories of food

based on a representation for food items that

calculates pairwise statistics between local features

has been presented (Yang et al., 2010). These statis-

tics are accumulated in a multi-dimensional histo-

gram, which is then used as input to a SVM classifier.

Food images are taken from the Pittsburgh Food Im-

age Dataset (PFID) (Chen et al., 2009). This system

has also achieved high levels of recognition accuracy

(80%).

The use of the BOF model has been adopted in

several food recognition systems recently, since food

recognition does not presume any typical spatial ar-

rangement of the food elements. Based on the BOF

model, the Food Intake Visual and Voice Recognizer

system which aims to measure the nutritional content

of a user’s meal (Puri et al., 2009) has been proposed.

Given a set of three images of a user’s plate of food,

the system first asks the user to list food items through

speech, then attempts to identify each food item on

the plate, and finally reconstructs them in 3D to meas-

ure their respective volumes. Food images are col-

lected by the developers of the system. Food classifi-

cation is based on the combined use of color neigh-

borhood and maximum response features in a texton

histogram model, which resembles the BOF ap-

proach. Adaboost is used for feature selection and

SVM for classification, which achieves recognition

accuracy about 90%. Moreover, a food recognition

system for the classification of Japanese food images

has been introduced (Joutou and Yanai, 2009), based

on the combined use of BOF of SIFT, Gabor filter re-

sponses and color histograms, which are then fed to a

multiple kernel learning classifier, which has

achieved acceptable levels of accuracy (61,34%). The

BOF model has been used in another automatic food

recognition system (Anthimopoulos et al., 2014). The

system firstly computes dense local features using the

SIFT on the HSV (Hue Saturation Value) color space,

then builds a visual vocabulary of 10000 visual words

by using the hierarchical k-means clustering and, fi-

nally, classifies the food images with a linear SVM

classifier, which achieves high levels of accuracy

(78%).

In the present study, a food recognition system is

proposed which consists of two modules performing

feature extraction and classification of food images,

respectively (Figure 1). Motivated by the ability of

SURF to capture spatial intensity patterns and the sta-

bility and distinctiveness provided by Color and Lo-

cal Binary Pattern (LBP) features, the combination of

SURF, Color and LBP features is examined in this

study. Moreover, a novel modified version of the All-

And-One (M-A&O) SVM classifier for multiclass

classification problems is proposed and its perfor-

mance is assessed against classification methods

A Modiﬁed All-and-One Classiﬁcation Algorithm Combined with the Bag-of-Features Model to Address the Food Recognition Task

285

based on SVM or the K-Nearest Neighbour ap-

proaches including the OAA SVM, the OAO SVM,

the A&O SVM, the Weighted K-Nearest Neighbour

(WKNN) classifier, the Dual Weighted K-Nearest

Neighbour (DWKNN) classifier, and the K-Nearest

Neighbour Equality (KNNE) classifier.

2 METHODS

2.1 Dataset

The Food Image Dataset (FID) used in this study con-

sists of 1200 images, 260-by-190 pixels each, col-

lected from the web. Each image belongs to one of six

categories corresponding to bread, meat, potatoes,

rice, pasta and vegetables (Figure 2). Each category

is represented by 200 images. The food is photo-

graphed under different servings, view angles, and

lighting conditions. The background of every image

is edited so that it is completely black.

2.2 Feature Extraction

In the present study SURF, Color and LBP features

have been combined to represent each food image in

the proposed food recognition system.

SURF detects points of interest using an integer

approximation of the determinant of Hessian blob de-

tector, and, then computes the features based on the

Haar wavelet response around each point of interest

(Bay et al., 2008). Color features are calculated as the

average value of color for every 4-by-4 pixel block of

the image. LBP is a texture descriptor that provides a

unified description, including both statistical and

structural characteristics of a texture patch (Prabhakar

and Praveen Kumar, 2012). The LBP feature vector

is calculated by dividing the image into cells, and

comparing the center pixel’s value with the neigh-

bours’ pixel values of each cell. Then, a histogram of

the numbers occurring over the cells is computed. A

useful extension to the LBP is the uniform LBP,

which reduces the length of the initial feature vector

from 256 to 59 (Ojala et al., 2002).

The approach of BOF is used to decrease the input

feature space, and deal with high visual diversity and

absence of spatial arrangement encountered in food

recognition. The BOF approach is influenced by the

Bag-Of-Words representation for text classification

(Cruz-Roa et al., 2011) and consists of the following

two steps. Firstly, a set of small blocks are extracted

from each image in the dataset, which are represented

by feature vectors. Secondly, the visual dictionary of

the image dataset is constructed and each image is

represented by the frequency of the codewords of the

visual dictionary. The visual dictionary is built with

the use of the k-means clustering algorithm. The clus-

ter centers of the feature points extracted in the first

step of the BOF approach are defined as visual words.

The visual dictionary is the combination of these vis-

ual words (Wang et al., 2016).

2.3 Classification

The classification task is performed using a modified

version of the All-and-One SVM and its performance

is assessed against several classification methods

based on the SVM and K-Nearest Neighbours (KNN)

approach, including the OAA SVM classifier, the

OAO SVM, the A&O SVM, the WKNN classifier,

the DWKNN classifier, and the KNNE classifier. All

algorithms have been implemented with MATLAB

2015a, are trained with the 70% of the images of the

FID, and tested with the rest 30% of the FID.

2.3.1 SVM-based Classifiers

The OAA SVM Algorithm.

The OAA SVM classifier (Galar et al., 2011) consists

of K binary SVM classifiers, where K is the total

number of classes. The i-th classifier is trained by la-

beling all the instances in the i-th class as positive and

the rest as negative. Each test instance is classified to

the class with the biggest score.

The OAO SVM Algorithm.

The OAO SVM classifier (Galar et al., 2011) consists

of K(K-1)/2 binary SVM classifiers, where K is the

number of classes. Each binary classifier learns to dis-

criminate between a pair of classes. The outputs of

these binary classifiers are combined so that the class

with the highest score is assigned to the test instance.

The A&O SVM Algorithm.

The A&O SVM algorithm (Pedrajas and Boyer,

2006) combines the strengths of the OAO and OAA

approaches. Taking into account that for a high pro-

portion of miss-classifications of the OAA approach,

the second best class is actually the correct class, and

that the binary classifiers of OAO are highly accurate

on their own, but may lead to incorrect results when

combined, the A&O approach combines the results of

K OAA classifiers and K(K-1)/2 OAO classifiers.

The A&O approach first classifies a test instance us-

ing the K OAA classifiers and holds the two classes

i,j with the biggest scores. Then, the binary classifier

of the OAO approach is used to classify the instance

among classes i,j.

HEALTHINF 2017 - 10th International Conference on Health Informatics

286

The M-A&O SVM Algorithm.

The M-A&O SVM algorithm combines the strengths

of the OAO and OAA approaches as the A&O SVM

algorithm, but in a different way. The M-A&O SVM

approach first classifies a test instance using the K

OAA SVM classifiers and holds the scores. Then, the

K(K-1)/2 SVM binary classifiers of the OAO ap-

proach are used to classify the instance. The test in-

stance will be assigned to the class that will achieve

the highest score from all (K + K(K-1)/2) classifiers.

2.3.2 KNN-based Classifiers

The WKNN Algorithm.

The WKNN algorithm is a modified version of the K-

Nearest Neighbours (KNN) algorithm. According to

the KNN algorithm, the k-nearest neighbours of the

query instance are selected according to a distance

criterion, such as the Euclidean distance. Then, the

query instance is assigned to the class represented by

the majority of its k-nearest neighbours in the training

set. In the WKNN algorithm, the closer neighbours

are weighed more heavily than the farther ones (Mari-

nakis et al., 2009) and the distance-weighted function





to the i-th nearest neighbor is defined as,





+1−

∑







where m is an integer in the interval (1,k) and k is the

total number of the neighbours.

The DWKNN Algorithm.

In order to address the effect of the number of neigh-

bours on the classification performance, a DWKNN

algorithm has been proposed (Gou et al., 2011). The

DWKNN algorithm gives different weights to the k

nearest neighbours depending on distances between

them and their ranking according to their distance

from the query object (Dalakleidi et al., 2013). The

distance-weighted function 



of the i-th nearest

neighbor is calculated according to the following

equation,





={







−











−







,





≠





1, 





=





where 





is the distance of the i-th nearest neigh-

bour from the query object, 





is the distance of the

nearest neighbour, and 





is the distance of the k-

furthest neighbour. Thus, the weight of the nearest

neighbor is 1, and the weight of the furthest k-th

neighbor is 0, whereas other weights are distributed

between 0 and 1.

The KNNE Algorithm.

The KNNE algorithm (Sierra et al., 2011) is a varia-

tion of the KNN classifier for multiclass classifica-

tion. It searches for the K-nearest neighbours in each

class and assigns the query instance in the class whose

K-nearest neighbours have the minimal mean dis-

tance to the test instance.

3 RESULTS

The FID is used for the evaluation of the proposed

classification algorithm against the classification al-

gorithms based on the SVM and KNN approach on

the food recognition task. In order to improve the

classification accuracy of the examined algorithms,

several sizes of the vocabularies of the BOF model

are tested. Table 1 shows the average accuracy of the

OAO SVM classifier on the six food classes for dif-

ferent sizes of the vocabulary of the BOF model for

SURF and Color features. The size of the vocabular-

ies has been varied from 100 to 2000 words. As it can

be observed from Table 1, the lowest accuracy (Acc

= 85.0%) is achieved with the size of 300 for both the

SURF and Color BOF vocabularies, whereas the

highest accuracy (Acc = 93.9%) is achieved with the

size of 1000 for both the SURF and Color BOF vo-

cabularies. It is also important to note that among the

three types of features, Color features contribute the

most to the accuracy of the OAO SVM classifier.

Table 1: Average accuracy of the OAO SVM classifier on

the six food classes of Food Image Dataset for varying size

of the vocabulary of the BOF model for SURF and Color

features.

Features

Acc

SURF Color LBP

100 100 59 87.5

200 200 59 90.0

300 300 59 85.0

400 400 59 90.6

500 500 59 91.1

600 600 59 91.7

700 700 59 92.5

800 800 59 93.1

900 900 59 90.6

1000 1000 59

93.9

1100 1100 59 93.3

1500 1500 59 91.4

2000 2000 59 90.8

A Modiﬁed All-and-One Classiﬁcation Algorithm Combined with the Bag-of-Features Model to Address the Food Recognition Task

287

Table 2: The average accuracy (%) of the classifiers under

comparison on the six food classes of the Food Image Da-

taset.

Algorithm Acc (%)

WKNN 84.4

DWKNN 92.8

KNNE 93.9

OAA SVM 90.6

OAO SVM 93.9

A&O SVM 90.3

M-A&O SVM

94.2

Table 3: Confusion matrix of the M-A&O SVM for each

food class (Bread, Meat, Pasta, Potatoes, Rice and Vegeta-

bles) of the Food Image Dataset.

Confusion Matrix

Acc

(%)

Br M Pa Pot R Veg

Br 93.3 0.0 0.0 0.0 6.7 0.0

M 1.7

95.

0.0 3.3 0.0 0.0

Pa 0.0 0.0 93.3 6.7 0.0 0.0

Pot 0.0 1.7 5.0 93.3 0.0 0.0

R 0.0 0.0 6.7 3.3

90.

0.0

Veg 0.0 0.0 0.0 0.0 0.0

100.

In Table 2, the average accuracy of the classifiers

under comparison on the six food classes is presented.

The size of the vocabulary of the BOF method for

SURF and Color features is 1000, thus a total number

of 2059 features is used for the classification. Ten k-

nearest neighbours are used for the WKNN, DWKNN

and KNNE. As it can be observed in Table 2, the low-

est average accuracy (Acc = 84.4%) is achieved by

the WKNN classifier, whereas the highest average ac-

curacy (Acc = 94.2%) is achieved by M-A&O SVM.

The second best average accuracy is achieved by the

OAO SVM and KNNE algorithms. The superiority of

M-A&O SVM can be explained by the fact that it

combines two very powerful strategies, the OAA

SVM and the OAO SVM, for multiclass classifica-

tion.

In Table 3, the classification accuracy of M-A&O

SVM for each food class is shown in the form of the

confusion matrix. It can be observed that the lower

classification accuracy (Acc = 90.0%) is achieved for

the class of rice. This is due to the fact that rice is

mingled with several sauces which can be very differ-

ent in color and texture. It is important to note that

rice is misclassified to potatoes and pasta which are

closer to it in terms of CHO than with meat or vege-

tables. The best classification accuracy is achieved

for vegetables (Acc = 100.0%), this is due to the dis-

tinctive green color of vegetables.

4 CONCLUSIONS

Automatic food recognition systems can be used to

estimate the content of a meal in CHO for patients

with diet related chronic diseases, such as obesity and

diabetes mellitus. In this study, an attempt to address

the tasks of feature extraction and food image recog-

nition was made. The use of the SURF, Color and

LBP features in combination with the BOF model has

proven to be particularly effective in terms of average

classification accuracy. Several classification ap-

proaches for multiclass classification have been

tested. The best classification accuracy (Acc =

94.2%) has been achieved by a modified version of

the All-And-One SVM approach and is quite high as

compared to reported values of classification accu-

racy for food images in the literature (60%-96%). The

proposed system can be combined with an image seg-

mentation module and a volume estimation module

towards the development of an automatic food recog-

nition system. Moreover, several other classifiers,

like AdaBoost, Random Forests and Convolutional

Neural Networks, can be used in the future for com-

parison purposes in the classification module.

Figure 1: Block diagram of the proposed system.

Figure 2: Example of images from each of the six categories

of the Food Image Dataset.

HEALTHINF 2017 - 10th International Conference on Health Informatics

288

ACKNOWLEDGEMENTS

The work of Kalliopi Dalakleidi was supported by a

scholarship for Ph.D. studies from the Hellenic State

Scholarships Foundation "IKY fellowships of excel-

lence for post-graduate studies in Greece-Siemens

Program".

REFERENCES

Anthimopoulos, M., Gianola, L., Scarnato, L., Diem, P. and

Mougiakakou, S., 2014. A Food Recognition System

for Diabetic Patients Based on an Optimized Bag-of-

Features Model. Journal of Biomedical and Health In-

formatics, 18(4), pp. 1261-1271.

Baumberg, A., 2000. Reliable feature matching across

widely separated views. In: IEEE Conference on Com-

puter Vision and Pattern Recognition. Hilton Head Is-

land, South Carolina, USA: IEEE, pp. 774 – 781.

Bay, H., Ess, A., Tuytelaars, T. and Gool, L., 2008.

Speeded-Up Robust Features (SURF). Computer Vi-

sion and Image Understanding, 110(3), pp. 346-359.

Carneiro, G. and Jepson, A., 2003. Multi-scale phase-based

local features. In: IEEE Conference on Computer Vi-

sion and Pattern Recognition. Madison, Wisconsin,

USA: IEEE, pp. 736 – 743.

Chen, M., Dhingra, K., Wu, W., Yang, L., Sukthankar, R.

and Yang, J., 2009. PFID: Pittsburgh fast-food image

dataset. In: IEEE International Conference on Image

Processing. Cairo, Egypt: IEEE, pp. 289-292.

Cruz-Roa, A., Caicedo, J. and Gonzalez, F., 2011. Visual

pattern mining in histology image collections using bag

of features. Artificial Intelligence in Medicine, 52, pp.

91-106.

Dalakleidi, K., Zarkogianni, K., Karamanos, V., Thanopou-

lou, A. and Nikita, K., 2013. A Hybrid Genetic Algo-

rithm for the Selection of the Critical Features for Risk

Prediction of Cardiovascular Complications in Type 2

Diabetes Patients. In: 13th International Conference on

BioInformatics and BioEngineering. Chania, Greece:

IEEE.

Florack, L., Haar Romeny, B., Koenderink, J. and

Viergever, M., 1994. General intensity transformations

and differential invariants. Journal of Mathematical Im-

aging and Vision, 4, pp. 171–187.

Freeman, W. and Adelson, E., 1991. The design and use of

steerable filters. IEEE Transactions on Pattern Analysis

and Machine Intelligence, 13, pp. 891 – 906.

Galar, M., Fernandez, A., Barrenechea, E., Bustince, H. and

Herrera, F., 2011. An overview of ensemble methods

for binary classifiers in multi-class problems: Experi-

mental study on one-vs-one and one-vs-all schemes.

Pattern Recognition, 44, pp. 1761-1776.

Gou, J., Xiong, T. and Kuang, Y., 2011. A novel weighted

voting for K-Nearest neighbour rule. Computers, 6(5),

pp. 833-840.

Joutou, T. and Yanai, K., 2009. A food image recognition

system with multiple kernel learning. In: IEEE Interna-

tional Conference on Image Processing. Cairo, Egypt:

IEEE, pp. 285–288.

Kong, F. and Tan, J., 2012. DietCam: Automatic dietary as-

sessment with mobile camera phones. Pervasive and

Mobile Computing, 8(1), pp. 147-163.

Marinakis, Y., Ntounias, G. and Jantzen, J., 2009. Pap

smear diagnosis using a hybrid intelligent scheme fo-

cusing on genetic algorithm based feature selection and

nearest neighbor classification. Computers in Biology

and Medicine, 39, pp. 69-78.

Martin, C., Kaya, S. and Gunturk, B., 2009. Quantification

of food intake using food image analysis. In: Engineer-

ing in Medicine and Biology Society Annual Interna-

tional Conference of the IEEE. Minneapolis, Minne-

sota, USA: IEEE, pp. 6869–6872.

Mikolajczyk, K. and Schmid, C., 2003. A performance

evaluation of local descriptors, In: IEEE Conference on

Computer Vision and Pattern Recognition. Madison,

Wisconsin, USA:IEEE, pp. 257 – 263.

Mindru, F., Tuytelaars, T., Van Gool, L. and Moons, T.,

2004. Moment invariants for recognition under chang-

ing viewpoint and illumination. Computer Vision and

Image Understanding, 94, pp. 3–27.

Ojala, T., Pietikäinen, M. and Mäenpää, T., 2002. Multi-

resolution Gray-scale and Rotation Invariant Texture

Classification with Local Binary Patterns. IEEE Trans-

actions on Pattern Analysis and Machine Intelligence,

24(7), pp. 971-987.

Pedrajas, N. and Boyer, D., 2006. Improving multiclass pat-

tern recognition by the combination of two strategies.

IEEE Transactions on Pattern Analysis and Machine

Intelligence, 28(6), pp. 1001–1006.

Peng, X., Wang, L., Wang, X. and Qiao, Y., 2016. Bag of

visual words and fusion methods for action recognition:

Comprehensive study and good practice. Computer Vi-

sion and Image Understanding, pp. 1-17.

Prabhakar, C. and Praveen Kumar, P., 2012. LBP-SURF

Descriptor with Color Invariant and Texture Based Fea-

tures for Underwater Images. In: 8th Indian Conference

on Computer Vision, Graphics and Image Processing.

Mumbai, India.

Puri, M., Zhu, Z., Yu, Q., Divakaran, A. and Sawhney, H.,

2009. Recognition and Volume Estimation of Food In-

take using a Mobile Device. In: Workshop on Applica-

tions of Computer Vision (WACV). Snowbird, Utah,

USA: IEEE, pp. 1-8.

Schaffalitzky, F. and Zisserman, A., 2002. Multi-view

matching for unordered image sets, or “How do I or-

ganize my holiday snaps?”. In: 7

European Confer-

ence on Computer Vision. Copenhagen, Denmark:

LNCS, pp. 414 – 431.

Shroff, G., Smailagic, A. and Siewiorek, D., 2008. Weara-

ble context-aware food recognition for calorie monitor-

ing. In: 12th IEEE International Symposium on Weara-

ble Computers. IEEE, pp. 119–120.

A Modiﬁed All-and-One Classiﬁcation Algorithm Combined with the Bag-of-Features Model to Address the Food Recognition Task

289

Sierra, B., Lazkano, E., Irigoien, I., Jauregi, E. and

Mendialdua, I., 2011. K nearest neighbor equality: giv-

ing equal chance to all existing classes. Information Sci-

ences, 181(23), pp. 5158–5168.

Wang, R., Ding, K., Yang, J. and Xue, L., 2016. A novel

method for image classification based on bag of visual

words. Journal of Visual Communication and Image

Representation, 40, pp. 24-33.

Yang, S., Chen, M., Pomerleau, D. and Sukthankar, R.,

2010. Food recognition using statistics of pairwise local

features. In: IEEE Conference on Computer Vision and

Pattern Recognition. San Francisco, California,

USA:IEEE, pp. 2249–2256.

Zhu, F., Bosch, M., Woo, I., Kim, S., Boushey, C., Ebert,

D. and Delp, E., 2010. The use of mobile devices in aid-

ing dietary assessment and evaluation. IEEE Journal of

Selected Topics in Signal Processing, 4(4), pp. 756–

766.

HEALTHINF 2017 - 10th International Conference on Health Informatics

290