Probabilistic Classifier for Fashion Image Grouping Using Multi-Layer Feature Extraction Model

Probabilistic Classifier for Fashion Image Grouping Using Multi-Layer Feature Extraction Model

Seema Wazarkar (National Institute of Technolgy Goa, Ponda, India), Bettahally N. Keshavamurthy (National Institute of Technology Goa, Ponda, India) and Ahsan Hussain (National Institute of Technology Goa, Ponda, India)
Copyright: © 2018 |Pages: 16
DOI: 10.4018/IJWSR.2018040105

Abstract

In this article, probabilistic classification model is designed for the fashion-related images collected from social networks. The proposed model is divided into two parts. The first is feature extraction where six important features are taken into consideration to deal with heterogeneous nature of the given images. The second classification is done with the help of probability computations to get collection of homogeneous images. Here, class-conditional probability of extracted features are calculated, then joint probability is used for the classification. Class label with maximum joint probability is assigned to the given image. A comparative study of proposed classification model with existing popular supervised as well as unsupervised classification approaches is done on the basis of obtained accuracy of the results. The effect of convolutional neural network inclusion in the proposed feature extraction model is also shown where it improves the accuracy of final results. The output of this system is useful further for fashion trend analysis.
Article Preview
Top

Introduction

In modern era, most people are conscious about their looks and current fashion trends. Everyone desires to be the first user of the upcoming fashion. Each fashion follows different periods of life cycle like short, moderate or long. Addressing fashion change in advance, helps in appropriate decision making at fashion industry. Popular events, social changes, economic conditions, subcultural influences, entertainment, technical innovations and fashion designers are some of the main reasons for fashion change. Hence, numerous factors need to be taken into account while fashion forecasting. It is a creative and repetitive process of future trend prediction. It follows a systematic procedure which includes information gathering, market research, consumer research and analysis (E. Kim et al., 2013).

Social content analysis (C. Aggarwal, 2011) is one of the methods used for fashion forecasting. Every day, huge amount of data gets uploaded on social networks by many users. To deal with these voluminous data, data mining techniques need to be applied (Nayak, 2010). Classification is a task of data mining which is used to assign a class label to the given data based on available labelled data. It is a supervised approach. If unlabelled data is used for this task, it is called as clustering, which is an unsupervised approach. In semi-supervised approach, small amount of labelled and large amount of unlabelled data is used. Social data is available in various forms such as numeric, text, audio, image, video, etc. For fashion forecasting, image is very useful and informative form of social content data. Initially, features of the given image need to be extracted and stored in vectors. Then, data mining task is performed on the obtained feature vectors.

Feature vectors store important representative information generated from a given image. Feature extraction is a generalized task and not a specific method. To accomplish this task various methods are available for each type of feature such as color, texture, edge, etc. Specific type of feature needs to be extracted in order to use it for a particular application. Obtained feature vector is appropriate, if it is able to provide good results for end user applications. For example, in case of “color forecasting for fashion”, color feature is important and useful to get good final results but other features like edges, textures, shapes are not useful in this case.

In this paper, semi-supervised model is proposed to extract image features and classify those images using joint probabilities of various features such as color, texture, regional, geometric, object detection (face) and number of maximum matching points using linear convolution. These features are taken into account to deal with heterogeneous nature of fashion images collected from social networks. Proposed model is referred as semi-supervised because both labelled and unlabelled set of images are used. Our contribution is given as follows:

  • Multi-Layer Feature Extraction Model (M-LFEM) is proposed, where multiple features such as color, texture and regional, geometric, face detection and number of maximum matching points are extracted to get feature vectors for a given image;

  • Given images are classified on the basis of joint probabilities of extracted features;

  • Performance of the proposed system is evaluated based on the accuracy of the obtained results.

Remaining paper is organized as follows: Related work is provided in Section 2. In Section 3, proposed work is presented. Then, experimental results are discussed in Section 4. Paper is concluded and future directions are provided in Section 5.

Top

Image data analysis contributes well in the area of fashion. Image classification provides an efficient way for image analysis. As proper pre-processing helps in getting accurate results, appropriate feature extraction is also an important part of the system. In this section, existing research works related to image feature extraction and fashion image classification are discussed.

Complete Article List

Search this Journal:
Reset
Open Access Articles
Volume 17: 4 Issues (2020): 1 Released, 3 Forthcoming
Volume 16: 4 Issues (2019)
Volume 15: 4 Issues (2018)
Volume 14: 4 Issues (2017)
Volume 13: 4 Issues (2016)
Volume 12: 4 Issues (2015)
Volume 11: 4 Issues (2014)
Volume 10: 4 Issues (2013)
Volume 9: 4 Issues (2012)
Volume 8: 4 Issues (2011)
Volume 7: 4 Issues (2010)
Volume 6: 4 Issues (2009)
Volume 5: 4 Issues (2008)
Volume 4: 4 Issues (2007)
Volume 3: 4 Issues (2006)
Volume 2: 4 Issues (2005)
Volume 1: 4 Issues (2004)
View Complete Journal Contents Listing