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

A Computer Vision System for Bangladeshi Local Mango Breed Detection Using Convolutional Neural Network (cnn) Models

A. S. M. Farhan Al Haque
Department of Computer Science and Engineering Daffodil International University

Md. Riazur Rahman
Department of Computer Science and Engineering Daffodil International University

Ahmed Al Marouf
Department of Computer Science and Engineering Daffodil International University

Md. Abbas Ali Khan
Department of Computer Science and Engineering Daffodil International University

Abstract/ Executive Summary:

Magnifera Indica, traditionally known as mango, is a drupe found around the world in over 500 species. India has produced 19.5 million metric tons of mango in 2017. In Bangladesh, mango has been referred as the national tree and government has included endemic species of mango as geographical index (GI) of Bangladesh. Recognizing specific breeds has become a significant computer vision task. In this paper, we have proposed the convolutional neural network (CNN) based approach for detecting five mango species namely, Chosha, Fazli, Harivanga, Lengra and Rupali from 15000 different images. For better experimentation, we have applied three different models of CNN and analyzed the recognition rates with various criteria. For performance evaluation, we have utilized the classic metrics such as precision, recall, F1-score, ROC and accuracy. Among the experimented three models, the third model, outperformed in terms of accuracy with 92.80%.

Keywords: Megnifera Indica, Mango Species Detection, Computer vision, Convolutional Neural Network (CNN)

Location: Department of Computer Science and Engineering Daffodil International University

Start Date:

End Date:

Program Area: Knowledge Management

Non-commodity: GIS

Objectives:

In this paper, we have formalized a RGB color image dataset of Bangladeshi local five mango breeds and utilized he dataset for detecting the mango breeds using ConvNet algorithm. For the experimental analysis, we have used three different ConvNet models with different parameters and reported the difference between them.

Methodology:

In this section, we have focused on the proposed model for our research to classify different fruits, their breeds and the quality of the fruits. The research work has been accumulated by the necessary steps described below in the subsection. A. Image Acquisition While the problem domain has been fixed and problem analysis has been completed the next phase of our work is to capture the images. Very few images were available in the internet and most of the images are captured by us in different locations and situations. A Nikon D7200 DSLR (24.2megapixel, ISO 100 - 25,600, liquid crystal display: 3.2 in. diagonal, 1,228,800 dots, wide viewing angle, thin film transistor based liquid crystal display, TFT-LCD screen, 6 frame per second shooting capacity, 24_16 mm image sensor, 51 point autofocus system) camera having 18-140 mm lens. In our research, we have captured images of 5 breeds of mango shown in Figure 2. B. Dataset To yield the best possible performance from our CNN model we have collected a prodigious dataset having 15000 images approximately 3000 images per sample class. The images are divided into training, validation and test set using the classical holdout method. The model has been trained with a considerable 2200 images for every sample class and validated by 400 images. Then the model was tested by another 400 different images to investigate the performance to classify. E C. Noise Removal One of the most key challenges to work with images is the noise which generally includes during the capture of images. As the performance of the CNN model mostly rely on the image quality, the noise exclusion techniques are very crucial. We have deployed a fuzzy filter to remove Gaussian noise. If input image is denoted by fp and fmax denotes the maximum intensity value among 8-neighboring pixels, then a function is calculated using equation 1. As we have used three RGB channels for CNN model, the filter works separately on the R, G and B components for noise reduction. The final result is obtained by concatenating the three component results using equation 1. D. Overfitting Elimination Overfitting is a major setback for the performance of any machine learning approaches which causes the model to memorize the details of the training data too closely but fail to extract feature. Due to the shortfall of generalizing the features the model cannot perform well on the test data. Resulting a very high accuracy for the training set and very poor result in the validation and test set. Several classical methods have been used to prevent this overfitting problem from the CNN model and boost accuracy. 1. Image Augmentation: Generally image collection and preprocessing phase is very time consuming and tedious. But equipping a deep learning model to produce great image classification results is not possible without a substantial number of training image. Image augmentation steps in with a viable solution creating different version of images from a training image by multiple operations of processing and manipulation. The API used in Keras for image augmentation is ImageDataGenerator. We have applied several techniques like re-scaling, shearing, height shifting, width shifting, zoom, rotation, horizontal flip etc. techniques are used to augment data. 2. Ridge Regression Regularization Simply L2 regularization reduces the complexity of the model by adding a squared magnitude co-efficient to penalize the given loss function. This least squared penalty term forces the model to under fit. It is proved to be a very efficient way to handle the overfitting problem. 3. Dropout Dropout is a probable solution for overfitting by dropping or ignoring a random number of neurons in the neural network. It is ideally efficient for introducing non linearity by forcing a deep learning model to learn more robust features and reduces dependencies on particular features or neurons. In our work, we have tried to turn off different number of neurons to get better results and experimentally found that 50% dropout generates the best accuracy result for classification of images. 4. Reduce Architecture Complexity Overtraining of the training data is very influential to cause overfitting. We have to be cautious about the number of convolutional layers, the number of filters deployed per layer, even the size and number of neurons in each layer of neural network. The selection of model configuration is very critical as we never want to apply unnecessarily complex model resulting overfitting.

Source of Information: 4th International Conference on Electrical Information and Communication Technology (EICT), 20-22 December 2019, Khulna, Bangladesh

Article Link (Online Journal):

Funding Source:

Budget:

Total Budget (Tk.) :

Achievement/Findings:

In our research, we have proposed a CNN based model that can classify among the different breeds of mango fruit with a very satisfactory performance. The result has been analyzed further with the performance metrics like precision, recall and F1 score to examine the rigorous performance and found to be performing comprehensively praiseworthy. The ROC curve has also exhibited a great shape with a very high AUCROC value of 97.3%. Later we have tried to show some optimization analysis for different learning rates, different value of dropout and varying number of epochs for execution. This work is still in progress. We are trying to add more breeds of mango and other fruits. We can apply more deep learning pretrained network like MobileNet, GoogleNet or AlexNet etc. to get even better results.

Knowledge Management: Report/Proceedings