Classification of Beverages Using Electronic Nose and Machine Vision Systems Mazlina Mamat * and Salina Abdul Samad † * Institute of Microengineering and Nanoelectronics, Universiti Kebangsaan Malaysia, Bangi Selangor, Malaysia E-mail: mazlina@eng.ukm.my Tel: +6-03-89216317 † Faculty of Engineering and Built Environment, Universiti Kebangsaan Malaysia, Bangi Selangor, Malaysia E-mail: salina@eng.ukm.my Tel: +6-03-89216317 Abstract— In this work, the classification of beverages was conducted using three approaches: by using the electronic nose alone, by using the machine vision alone and by using the combination of electronic nose and machine vision. A total of two hundred and twenty eight beverages from fifteen different brands were used in this classification problem. A supervised Support Vector Machine was used to classify beverages according to their brands. Results show that by using the electronic nose alone and the machine vision alone were able to respectively classify 73.7% and 92.9% of the beverages correctly. When combining the electronic nose and the machine vision, the classification accuracy increased to 96.6%. Based on the results, it can be concluded that the combination of the electronic nose and machine vision is able to extract more information from the sample, hence improving the classification accuracy. I. INTRODUCTION Since its emergence around 30 years ago, the electronic nose technology has been advancing tremendously. From a simple system that was capable to solve simple odor based classification problem [1], it is now equipped with sophisticated accessories and has been used to solve many problems involving odor analysis in various industries [2]. Many literatures reported the application of electronic nose in the food industry, where it was used to evaluate food freshness [3], to determine storage stability of food product [4], to determine the fruit ripeness level [5], to determine food shelf- life [6], to distinguish different types of beverages [7], to name a few. The extensive applications of electronic nose in food industry were due to the nature of food that emits unique odor which can be fairly distinguished. Typically, the electronic nose operates as follows: detect the volatiles emitted by the food sample by using the gas sensor array, obtain the valuable information from each of the sensor response, and finally identify the food sample. To execute the above operation, the electronic nose has a gas sensor array, a data acquisition and controller system and a data analysis software. The gas sensor array consists of several chemical gas sensors with different selectivity and sensitivity. The number, type and selectivity of the sensor array are determined either using the blackboard approach [8] or by using the experimental approach [2]. On average, 3 to 8 sensors were normally used, however the number can go to 32 sensors [9]. Each sensor in the array interacts with the volatiles emitted by the food sample. The interaction produces certain changes in the sensor, usually presented by the increasing or decreasing of resistance. The resistance changes were further presented as voltage difference that can be processed by the data acquisition and controller system. The information contain in the sensor response was extracted. The combination of information obtain from every sensor response will form a pattern that represents the food sample. Based on this pattern, the data analysis software will identify the food sample. The electronic nose alone was proven able to solve many odor based classification problems with high accuracy [10, 6]. However, efforts were made to improve the classification performance especially for difficult problems. Among the efforts are: using different sensors technology, using different features of the sensor response, and using different data analysis techniques [11]. Apart from modifying the electronic nose itself, there were attempts to combine the electronic nose with other tools such as the electronic tongue [12, 13], the mass spectrometer [14] and the machine vision [15, 16]. These combinations were performed to extract more information from the sample in order to improve discrimination between classes. However the electronic tongue requires direct contact to the sample while the mass spectrometer requires the sample to be vaporized. These requirements spoil the contents and destroy the sample physically, thus are considered destructive. The machine vision on the other hand offers the ability to extract physical information of the sample while preserving the sample. This paper presents the classification of beverages by using the combination of electronic nose and machine vision systems. As comparison, the classification of beverages was also performed by using the electronic nose system individually and machine vision individually. The electronic nose and machine vision systems were described briefly in the following sections. The classification of beverages was executed with the help of supervised Support Vector Machine.