CHAPTER I

PROBLEM AND ITS BACKGROUND

This chapter includes the introduction, objectives, significance of the study,

conceptual framework, scope and delimitation and the definition of terms to be used.

1.1 Introduction

Despite its geographical advantage, Philippines face large demand in coffee

production. Typical coffees that grow in the Philippines are Robusta, Arabica, Liberica

and Excelsa. Batangas province gained the name “Coffee Capital of the Philippines” for

producing great and unique type of coffee, the “Kapeng Barako” or“ Coffea Liberica”, the

glory days of Batangas came to an end when the disease coffee leaf rust hit the

plantations in the late 90’s(Cantos, 2013).

The Philippines became a member of the International Coffee Organization (ICO)

in 1980 wherein a group of importing and exporting coffee nations(De Guzman, 2017).

According to Philippine Statistics Authority (PSA) industrial crops quarterly report from

January to March 2018, coffee production went down by 8.7 % this quarter from 20.19

thousand metric tons to its current level of 18.42 thousand metric tons for the same

period last year. Based on the PSA 2018 quarterly report, the most produced variety was

Robusta coffee at 72.5 %, followed by Arabica and Excelsa at 18.4 % and 8.5 %,

respectively while only 0.6 % of the national coffee production was produced by Liberica

variety(Major Non-Food and Industrial Crops Quarterly Bulletin, January-March 2018,

2018). According to Asian Institute of Management, small farmers did not found coffee

farming attractive due to the low income and high cost of producing the crops despite

high demand for the Liberica coffee (Padin, 2016).

1
 Before industries achieved the perfectly brewed coffee, it undergoes different

process from harvesting up to the sorting of quality coffee green beans. Sorting process

is an essential stage to improve coffee quality. Traditionally, sorting process is done by

hand and anything that looks too dark or damaged is removed. Manual sorting of coffee

green beans consumes more time and also causes inconsistency in grading that leads

to slow production in coffee industry.

Katy’s farm is one of the biggest producers of Barako coffee in Lipa City, wherein

they experience a difficulty in sorting green coffee beans. According to the farm

manager, 20% of the produced green coffee beans after removing the parchment were

qualified for exporting/importing and the remaining 80% of the beans were not classified

as total rejects, but can later be market in the local.

In response to this, several studies were already conducted to develop a

technology that can provide an automated system specifically in agriculture with the help

of advance computers. Whereas, computer-aided applications already created a big

impact in the area of agriculture to aid farmers to work efficiently, a particular process

that uses advance artificial intelligence algorithms. Artificial intelligence is already paving

the future of agriculture and emerging in terms of automation, soil and crop monitoring

and predictive analysis.

Relating to this research study entitled with “Green Beans Classifier for

Liberica Coffee using Deep Convolutional Neural Network”, a Deep Neural Network

classifier will be developed that will capture coffee green beans and detects the physical

quality according to its class. The researchers believes that this study would significantly

improve the traditional manual sorting of coffee, to address the problem a convolutional

neural network classifier would be developed to properly maintain the consistency in

grading, uniformity, and to impose time delay of coffee production.

2
1.2 Objectives

This research study aims to develop and implement a deep convolutional neural

network classifier to detect beans with physical defects. Specifically,

1. Preparation of image dataset and identify the different type of defects in the

green coffee beans.

2. Train the Deep Neural Network using the acquired green coffee beans

dataset through transfer of learning.

3. Implement the trained model using computer-based classifier.

4. Test and evaluate the implemented deep convolutional neural network

classifier.

1.3 Conceptual Framework

The conceptual framework presented and discussed in the following summarizes

the variables involved in developing the project and how they were put together as a

whole system.

Figure 1 shows the conceptualized ideas in order to achieve the main objectives

of the research.

3
 Figure 1 Conceptual Paradigm

The first phase of this study was to determine the concepts of Deep Neural

Network, and other useful attributes related to the development of the classifier. Based

on available references and related works, the necessary tools needed to build the

classifier have been determined. This also includes software programming for the

development of a deep neural network model, data acquisition, and image processing

tools to identify the class of green coffee beans.

After the data and conceptual gathering, the development stage has been

followed wherein the researchers conducted selection of state-of-the-art CNN models,

training of acquired models through the green coffee beans dataset, and evaluation of

the trained models to measure its accuracies to the desired output.

4
 The final phase of this research study is to implement the model with the highest

accuracy. It includes testing and troubleshooting if there was any problem encountered

with the classifier. Finally, a script was developed to visualize the effectiveness of a

neural network in a computer screen.

1.4 Significance of the Study

Coffee quality is of vital importance to the coffee industry. In traditional practice of

manual sorting, color and coffee bean size are considered as a key indicator of coffee

quality for all varieties and a particular importance to those who want to meet beyond

minimum export standards and toward a higher level of product. Manual operation of

sorting green coffee beans is considered one of the factors affecting the low production

of coffee. One way to address this problem is through the implementation of a computer

based system whose input data is digital images of green coffee beans acquired using a

high resolution camera.

The system is also developed in a way that the process is continuous.

Furthermore, green coffee beans are captured in a robust setting since the system

needed a large amount of training data in order to achieve a better performance. In

classifying, the system will determine if the captured green coffee beans were damaged

or not. Listed below are the importance and benefits of the study.

 To other researchers who want to pursue automation in agriculture, this neural

network classifier can be used as a starting point to implement a machine sorting

system wherein it will automatically detect the quality of the crops and sort it

according to its class.

5
  To other developers that are planning to go beyond this field of technology, the

Liberica raw beans dataset generated by the researchers can be used in building

their own classifier.

 To the farm owners and farmers, the computer-aided classifier system will be

developed not specifically to stop the manual operation of sorting coffee green

beans, instead to assist the farmers in classifying the quality beans accurately.

The importance of the system is to improve the traditional manual sorting of

green coffee beans. Through this deep neural network classifier, the possibilities

of error will reduce.

1.5 Scope and Delimitation of the Study

The main focus of this study was to develop a computer-based deep neural

network classifier that detects and recognizes coffee green beans based on physical

quality appearance specifically the broken, sour, pea berry and blacked beans.

The study was exclusive only for the green beans of Liberica variety of coffee.

The green coffee beans were captured in different height, angles, backgrounds and

lighting conditions, since there was no standard procedure in taking image samples. A

Logitech HD C525 webcam and Sony Xperia Z5 mobile phone was the two camera

devices used in taking coffee green beans samples.

Transfer of learning was used through a pre-trained CNN model such as Faster

RCNN with Inception V2, Faster RCNN with Resnet50, SSD with Mobilenet V1, SSD

with Mobilenet V2 and SSD with Inception V2 as a starting point to train data rather than

creating a new one and trained from scratch. The researchers selected the state-of-the-

art pre-trained CNN models because it already achieved a high accuracy and trained

over a thousand of input images. The researchers used the Tensorflow Object Detection

API which consists of compilation of python scripts that can be used for training the

6
neural network. The deep neural network classifier has been trained with an Intel Core i7

Processor and Nvidia Geforce 1050 GPU.

Further research regarding this study, the system can be used as a starting point

to implement a machine sorting system. The researchers chose to focus on a computer-

based system since creating a machine is another field which requires additional

knowledge that covers mechanism.

1.6 Definition of Terms

Algorithm - a systematic procedure designed to solve a problem or perform a specific

task (Merriam Webster Dictionary, n.d.)

Classifier –a computer based system made using transfer learning through the use of

pre-trained CNN model to classify green coffee beans

Convolutional Neural Network (CNN) - a deep, feed-forward neural network that has

been applied to analyse visual imagery (Github, n.d.)

Deep Learning – machine learning subfield dealing with algorithms stimulated by the

structure and function of the brain. (Brownlee, 2016)

Deep Neural Network (DNN) - consists of multiple layers that are hidden within the

input layers and output layers of an artificial neural network (Deep Learning, n.d.)

Graphical User Interface –allows the use of symbols or icons to interact with the

user(Encyclopedia Britannica, n.d.)

Coffee Green Beans –the unroasted version of coffee beans(About Green Coffee

Beans FAQ, n.d.)

Transfer of Learning - a method in machine learning whereas the developed model for

a given task is reused of the second task as the starting point of the model (Brownlee,

2017)

7
Tensorboard – tool in deep learning wherein it manages and shows the data achieved

from the start of the training until the end

Region-of-Interest- part of an image that is being filtered and created a binary mask

image (MathWorks, 2018)

Precision – it measures the classifier prediction on the ratio of true objects to the total

number of objects or the “false positive rate” (Arlen, n.d.)

Recall –it measures “false negative rate” or the ratio of true object detections to the total

number of objects in the data set (Arlen, n.d.)

Pascal VOC format - is an xml file that is used by Imagenet (Enrique, 2018)