VISION-BASED ALGORITHM FOR RECYCLABLE WASTE
CLASSIFICATION
Aaron Don M. Africa, Neil Patrick J. Labay, Mariel Sheryn B. Ong, Jillian Abigail J. Sales and
Mitsuru Mike E. Toyoda
Department of Electronics and Communications Engineering, Gokongwei College of Engineering, De La Salle University, Manila, Philippines
E-Mail: [email protected]
ABSTRACT
Waste segregation is one of the most prominent problem in the worl particularly in the Philippines where waste segregation is done manually. This research aims to solve the problem regarding waste segregation by means of image processing. These recyclable materials are classified into nine groups namely: aerosol cans, aluminum cans, cereal box, glass bottles, paper bowls, plastic bottles, plastic cups, tetra packs and tin cans. The recyclable material is subjected to a controlled environment then the image is captured by the camera. The region of interest is extracted from the image by the use of cascade filters such as weiner and median filter along with morphological operators and canny edge detecto. Scale Invariant Feature Transform (SIFT) features are extracted from the image. Then two methods are used for classfying, namely: Random Sampling and Consensus (RanSac) and a combination of Bag-of-Words (BOW) and Support Vector Machines (SVM).
Keywords: waste segregation, waste classification, canny edge detection, median filter.
1. INTRODUCTION
Waste management is one the biggest problem that the world is facing. As a matter of fact, according to a 2012 World Bank report, global solid waste generation is set to increase 70 percent by 2025, rising from more than 3.5 million tons per day in 2010 to more than 6 million tons per day by 2025 [1]. It is also estimated that the global cost would also increase from $205 billion a year in 2010 to $375 billion by 2025[2]. In the case of the Philippines, it is much worse. Some people just throw their garbage anywhere from streets, to rivers, etc. They are not aware of the hazard that it entails. According to a study, there are about one thousand and five hundred (1,500) tons of garbage that are dumped illegally everyday in private lands and bodies of water while others are burned openly causing air pollution [3]. A solution for this is to employ waste segregation to reduce the amount of waste. However, there is no clear indication of waste segregation in the Philippines.
This research aims to solve the problem in waste segregation. A similar research was also done to solve this problem. However, by using mechanical hardware and sensors to segregate recyclable materials tend to have a lower range of waste classification. It also entails larger overhead cost because of the hardware and is susceptible to errors. Thus the group implemented it using image processing where there is a lower overhead cost. This is because only a camera is needed to classify the image. Furthermore, with the use of image processing algorithms and statistical analysis, it would not be susceptible to errors.
In this research, the recyclable materials will be classified into nine (9) categories. These are aerosol cans, aluminum cans, cereal boxes, glass bottles, paper bowls, plastic bottles, plastic cups, tetra packs and tin cans. The group used two methods in classifying the images. These are RanSac and BOW - SVM.
2. IMAGE ACQUISITION
Prior to the image acquisition part, it was to be assumed that the waste materials had undergone a community-based segregation. This means that the waste was already been segregated from recyclables to non-recyclable waste materials. These images have unknown data. For the missing data a Rough Set Theory [4, 5] based algorithm was used for testing [6, 7].
During the image acquisition part, each recyclable material was put into a picture box as shown in Figure-1. This picture box simulates a controlled environment which means that the image was set at a fixed lighting condition and captured by the camera at a fixed distance. A simulation of a fixed environment is done to ensure the consistency of each image to be taken and to prevent other distortions and noises entailing an uncontrolled environment [8]. A black background was used to ensure that transparent wastes such as plastic bottles and plastic cups can be easily identified through the contrast of colors. Another important factor that should be considered is the communication of the system [9, 10]. The networking side of things should be considered [11].
The camera used was a digital single-lens reflex camera (DSLR) with a five megapixel (5MP) configuration. This was used to ensure that the image acquired was of a decent quality.
3. IMAGE PRE-PROCESSING
The image pre-processing part was comprised of three parts. These were constructing an image database, image filtering and ROI extraction. For the image extraction Logic Scoring of Preference (LSP) can be an alternative reliability scheme for checking [12].
3.1 Image database
After capturing the image, it is stored in an image database which is a file folder located in the local disk. The image database contains one hundred and fifty images of each classification at different locations for the test set. Forty images of each classification at a fixed location for the training set are stored.
3.2 Image filtering
As the image is loaded into the MATLAB, it was subjected into different image filters as shown in Figure-2 [13]. First, the image is resized into a 768 x 1024 resolution to easily process it. Then the image was converted into an unsigned eight-bit (8-bit) grayscale image to easily process it. Then, a Canny edge detector was used to extract the edges of the image given its built-in Gaussian filter and ability to identify more edges that the other edge detector [14]. Next, the image was subjected to a wiener filter to reduce the noise in the image. Then, the median filter was used to smooth the edges in the image. Then, a filling algorithm was used to fill in holes in the image. The image was complimented which means that the color was reversed to ensure that only the silhouette of the object was seen. Then, morphological operators, which are practical methods to eliminate unwanted pixels near the ROI [15], were used remove the other noise surrounding the image. Artificial Neural Network [16] and Spatial Imaging [17] can also be used to filter the systems.
Figure-2. Raw Image (row 1: left); Grayscale Image (row 1: right); Canny Edge Image (row 2: left); Wiener Filter Image (row 2: right); Median Filter Image (row 3: left); Fill Image (row 3: right); Compliment Image (row 4: left); Silhouette Image (row 4: right).
3.3 ROI extraction
As the image silhouette was obtained, this was subjected to regionprops function to identify the properties of an image. The area of the binary image was identified along its bounding box. As the bounding box was identified, it is extracted from the whole image. The final product is shown in Figure-3. In the ROI extraction a database configuration is needed. The configuration is patterned using the researches from [18, 19, 20].
Figure-3. Raw Image (left); ROI extracted image (right).
4. IMAGE CLASSIFICATION
4.1 Random sampling and consensus method
4.2 BOW-SVM method
Just like the first method, SIFT features were extracted from each image in the training database. Then, all of the SIFT features of all of the training images are collected. By using the divide and conquer approach, the data can be easily analyzed. By the use of K-Means clustering, the set of points shall be clustered or grouped so that they be can easily analyzed [23]. Next, using the Bag-of-Words (BOW) algorithm, the key points are further classified. Also with the use of BOW algorithm, the system creates a dictionary of words. Each word is an identifier of each classification [24]. Finally, with the use of multiple SVM, the images can be trained to identify each set of images. This method was adapted from a previous study made by Chyi-Yeu Lin in “Object orientation recognition based on SIFT and SVM by using
stereo camera [25].”
The SVM configuration was RBF kernel SVM with a Dictionary Size set at 1500 for 400 training data set. This method requires proper communication of Computer Networks to work [26, 27]. This includes proper network configuration [28].
5. DATA AND RESULTS
5.1 Random sampling and consensus method
Using ten (10) test runs with a total of two hundred (200) images in each run, this yielded the following results. The distribution of each item was set at random.
Table-1. RanSac Test Results (Test 1 - Test 5).
Table-2. RanSac Test Results (Test 6 - Test 10).
The RanSac method was used to match the features extracted from the test image to one of the random samples in the training data. The overall recognition per run shows that the algorithm works at an average of 83.8%. However, the categorical tests showed
that the algorithm becomes inconsistent because of the assignment of threshold values placed per test run of the program which is difficult when variety and long operating time are considered.
5.2. BOW-SVM method
Table-3. BOW - SVM Test Results (Test 1 - Test 5).
Table-4. BOW - SVM Test Results (Test 6 - Test 10).
The BOW - SVM is most efficient algorithm used by the group is the method composing of bag-of-words and support vector machine. The features extracted using the Scale-Invariant Feature Transform will be treated as a visual word by the use of k-means in order to cluster these extracted features and to form a class or a group. The clustered group features are then passed on to the support vector machine classifier. In this third method, the group was able to come up with an algorithm with an overall accuracy of 97.19 percent for assorted items in a test set.
6. CONCLUSIONS AND RECOMMENDATIONS It has been shown that for the image classification tasks that the Bag-of-Words - SVM method has the highest recognition rate for a high-variety classification problem. Other image feature extraction techniques such as SURF, Dense SIFT, and PHOW may be used with Bag-of-Features and SVM for classification and produce the same satisfactory results. The visual variety of objects is also taken into account; therefore it is much likely to retrain the system to anticipate objects that are more common to be subjected into the system and is expected to be recognized correctly. Therefore, the Bag-of-Features-SVM recognition method works particularly well given that the background scene is constant and the training data used for the model is much likely to be anticipated to be subjected to the system.
for classifying heterogeneous objects placed in the belt as a feedback mechanism. In this type of system, it is not always the case that a single object can be placed within the captured area. Therefore, the multiple image detection can also be used for the image processing module of the system.
On the other aspects of the classifier, it is highly recommended that the system be deployed in a distributed manner, in which the distributed server will take care of the recognition process and the clients will take care of the image acquisition, feature extraction and data transformation, and the final action based on the decision made by the server. This can be done by using several connections to the database and the computing server. It is also recommended that further image processing be made for the detected image before extracting its features to decrease the variance of the image properties such as illumination, or reduce the dimensionality of the feature vectors before subjecting to the recognition process. The machine learning of the classifier is currently active for research and is highly recommended for further studies in artificial intelligence. The group recommends a study in which if the classifier does not recognize the object properly, the system must be capable of learning the type of object and store it into its knowledge base automatically.
REFERENCES
[1] Global Waste on Pace to Triple by 2100. 2014. http://www.worldbank.org/en/news/feature/2013/10/3 0/global-waste-on-pace-to-triple
[2] Hoomweg D., et al. 2014. Environment: Waste production must peak this century.http://www.nature.com/news/environment-waste-production-must-peak-this-century-1.14032
[3] Mair D. A., et al. 2014. Solid waste management in the Philippines: a small island experience.http://www.envirocorp.com/img/media/S WM_Philippines_Paper.pdf
[4] Africa A. 2017. A Rough Set-Based Expert System for diagnosing information system communication networks. International Journal of Information and Communication Technology. 11(4): 496-512.
[5] Africa A. and Cabatuan, M. 2015. A Rough Set Based Data Model for Breast Cancer Mammographic Mass Diagnostics. International Journal of Biomedical Engineering and Technology. 18(4): 359-369.
[6] Africa A. 2017. A Rough Set Based Solar Powered Flood Water Purification System with a Fuzzy Logic Model. ARPN Journal of Engineering and Applied Sciences. 12(3): 638-647.
[7] Africa A. 2017. A Mathematical Fuzzy Logic Control Systems Model Using Rough Set Theory for Robot Applications. Journal of Telecommunication, Electronic and Computer Engineering. 9(2-8): 7-11
[8] Picon A., et al. 2010. Automation of waste recycling using hyperspectral image analysis. IEEE Conference on, Bilbao. 1-5.
[9] Africa A., Chua A., Chua S., Gregorio J. and Quilnet M. 2019. Two-way SMS and braille communication for the visually impaired. ARPN Journal of Engineering and Applied Sciences. 14(2): 425-432.
[10]Sahin M., Taskiran, Z., Guler H., and Hamaci S. 2019. Simulation and Implementation of Memristive Chaotic System and its application for communication systems. Sensors and Actuators A: Physical. 290(1): 257-266.
[11]Zhang R., Wang X., Cheng M., and Jia T. 2019. The evolution of network controllability in growing networks. Physica A: Statistical Mechanics and its Applications. 520(1): 188-191.
[12]Africa A. 2018. A Logic Scoring of Preference Algorithm using ISO/IEC 25010:2011 for Open Source Web Applications Moodle and Wordpress. ARPN Journal of Engineering and Applied Sciences. 13(15).
[13]Brucal S., Africa A. and Dadios E. 2018. Female Voice Recognition using Artificial Neural Networks and MATLAB Voicebox Toolbox. Journal of Telecommunication, Electronic and Computer Engineering. 10(1-4): 133-138.
[14]Canny J. 1986. A computational approach to edge detection. IEEE Transactions on Pattern analysis and Machine Intelligence. 8(6): 679-698.
[15]Kachouri R., et al. 2008. Feature extraction and relevance evaluation for heterogeneous image database recognition. Image Processing Theory, Tools and Applications. 1-6.
[16]Africa A. and Velasco J. 2017. Development of a Urine Strip Analyzer using Artificial Neural Network using an Android Phone. ARPN Journal of Engineering and Applied Sciences. 12(6): 1706-1712.
Telecommunication, Electronic and Computer Engineering. 10(1-5): 15-20.
[18]Africa A., Aguilar J., Lim Jr., C., Pacheco P. and Rodrin S. 2017. Automated Aquaculture System that Regulates Ph, Temperature and Ammonia. 9th International Conference on Humanoid, Nanotechnology, Information Technology, Communication and Control, Environment, and Management (HNICEM).
[19]Africa A., Mesina A., Izon J. and Quitevis B. 2017. Development of a Novel Android Controlled USB File Transfer Hub. Journal of Telecommunication, Electronic and Computer Engineering. 9(2-8): 1-5.
[20]Saraireh S. 2019. Providing Confidentiality, Data Integrity and Authentication of Transmitted Information. ARPN Journal of Engineering and Applied Sciences. 14(1): 278-285.
[21]Lowe D. G. 2004. Distinctive image features from scale-invariant keypoints. http://www.cs.ubc.ca/~lowe/papers/ijcv04.pdf
[22]He L., et al. 2011. Object detection by parts using appearance, structural and shape features. International Conference on, Beijing. 489-494.
[23]Schaefer G. 2010. Content-based retrieval from image databases: Colour, compression, and browsing. International Conference onShah Alam, Selangor. 5-10.
[24]Csurka G., et al. 2014. Visual categorization with Bag
of Keypoints.
http://www.cs.cmu.edu/~efros/courses/LBMV07/Pape rs/csurka-eccv-04.pdf
[25]Chyi-Yeu L. and Setiawan E. 2008. Object orientation recognition based on SIFT and SVM by using stereo camera. IEEE International Conference on, Bangkok. 1371-1376.
[26]Guanzon J., Caberte C. and Dela Rosa J. 2019. An Expert System Optimization Model for Desktop Computers. International Journal of Advance Trends in Computer Science and Engineering (IJATCSE). 8(3): 676-682.
[27]Sayoc V., Dolores T., Lim M., and Sam Miguel L. 2019. Computer Systems in Analytical Applications. International Journal of Advance Trends in Computer Science and Engineering (IJATCSE). 8(3): 772-777.
