Agricultural Research Management Information System

Home
Research Summary
Search
About Us
- ARMIS
- Brochure
Contact Us
Report
User Request
Data Input
Help
- Operation Manual
  - PDF
  - Video
- Program Area & Commodity
We have reached 37600 number of research entries at this moment.

- Logout

Research Detail

Home
Research
Detail

Quality Inspection, Maturity Detection, and Size Based Grading of Various Types of Mangoes Using Machine Learning Methods

Farhana Tazmim Pinki*
Student
Computer Science and Engineering Discipline, Khulna University, Khulna, Bangladesh

S.M. Mohidul Islam
Associate Professor
Computer Science and Engineering Discipline, Khulna University, Khulna, Bangladesh

Abstract/ Executive Summary:

Mango is the most significant and flavorsome fruit in most continents, especially in Asia. Mango grading is an important task in the agro-industry. The grading process creates many problems during harvesting for mango growers. The manual grading process is performed by visual inspection and it is very time-consuming and labor-intensive. Due to different market prices and different market demands, automation in mango grading plays an important role to achieve better accuracy and consistency. In this study, an automatic mango grading system is developed using machine learning and image processing techniques. The system is divided into four phases. In the first phase, quality inspection of mango is performed using Convolution Neural Network (CNN) to detect healthy and diseased mango. In the second phase, different types of healthy mangoes such as Badami, Kesar, and Totapuri are classified using the ensemble method, Random forest. In the third phase, maturity detection is performed using another ensemble method, AdaBoost for the specific type of healthy mangoes to detect ripe, unripe, and partially ripe mangoes. Finally, in the fourth phase, size-based grading is performed on the specific type and maturity to determine large, medium, and small mangoes using K-nearest neighbor. Thus the different grades of mangoes based on quality, type, maturity, and size are obtained which have different market prices and demands. From experiments, the system shows 94.52% average accuracy.

Keywords: Quality inspection, Classification, Maturity detection, Size based grading, Feature vector, Machine learning methods

Location: Computer Science and Engineering Discipline, Khulna University, Khulna, Bangladesh

Start Date:

End Date:

Program Area: Farm Mechanization

Commodity: Mango

Objectives:

The objective of this research is automatic grading of various types of mangoes with specific quality, maturity, and size. At first, we use Convolution Neural Network (CNN) for quality inspection of mango.

Methodology:

In the proposed system, the quality of mangoes is inspected first. This is called quality inspection. From this quality inspection, we separate the healthy and diseased mangoes. Then only healthy mangoes are classified into various types. We call it the mango classification step. After classification, we detect the maturity of each type of healthy mangoes. Thus, we get ripe or partially ripe or unripe mangoes. This is the maturity detection step. After that, the healthy mangoes with a specific type of specific maturity are categorized based on their size and we get large or medium or small-sized mangoes. This is called size based grading step. After completing all of the processes, finally, we get the healthy mangoes of a specific type with specific maturity for a specific size. This automatic grading on various perspectives will be helpful for mango manufacturing after harvesting. Quality Inspection The system architecture of the quality inspection step is described below. Image Acquisition The mango images are collected from the field or the internet for quality-based grading. Image Preprocessing The captured images are in RGB color space which is a model of color images. Since RGB color space depends on device, it is mapped to grayscale image. Then the images are resized to make all a same size image. The images may contain unnecessary detail and insufficient brightness while collecting from field. So to remove noise, median filter is applied. The images may also have an unnecessary background which can affect while extracting features. To separate the mango from the background, the background of image is removed using the background removal technique. Quality Inspection Using CNN The image acquisition and preprocessing steps are similar to both training and testing phases. CNN is very similar to regular neural networks. Since it is very complex model and consists of up to hundreds of millions of parameters, CNNs need large training datasets to achieve accurate results [7]. We have used three convolution layers and three hidden layers in our convolution neural network architecture. The input shape is 64×64. It is sent to the convolution layer that has 32 feature maps and 3×3 filter size. We use Rectified Linear Unit (ReLU) activation function. It only passes the positive value or zero. Then we send it to another two convolution layers that have 64 and 128 feature maps respectively and the filter size is 3×3. The feature map is converted into a flattened array after finishing the convolution layer. Then it is sent to the hidden layer which is fully connected. We use 128 neurons in all three hidden layers with the activation function ReLU. The output layer with a sigmoid function is fully connected with the previously hidden layer. We use a sigmoid function which outputs the value between 0 and 1. As our model is a binary classifier and output is 0 or 1 like a predictive model, we use the sigmoid function here.

Source of Information: Journal of Image Processing & Pattern Recognition Progress Volume 7, Issue 1 ISSN: 2394-1995

Article Link (Online Journal):

Funding Source:

Budget:

Total Budget (Tk.) :

Achievement/Findings:

In this research, the automatic mango grading system is developed based on quality, type, maturity, and size using machine learning methods. For quality inspection, the CNN model is used to detect the quality of mango whether it is healthy or not. Then the healthy mangoes are classified according to their types by extracting mean, mode, and standard deviation values and using a random forest classifier. To perform maturity detection, color moment and color histogram values are extracted and the AdaBoost classifier is used to detect ripeness of mangoes. Finally, size feature such as the area is extracted and Hough transform is performed on the specific type and maturity based mangoes and KNN classifier is used to detect large, medium, and small-sized mangoes. Thus an automatic mango grading system is developed which will help mango growers and other related persons during harvesting and marketing in both rural and urban areas. The experimental results demonstrate that the overall system accuracy outperforms many existing systems.

Knowledge Management: Journal