Instance Segmentation and Classification of Coffee Leaf Plant using Mask RCNN and Transfer Learning | ||||
| Fayoum University Journal of Engineering | ||||
| Volume 7, Issue 1, January 2024, Page 130-141 PDF (736.16 K) | ||||
| Document Type: Original Article | ||||
| DOI: 10.21608/fuje.2023.226247.1057 | ||||
 View on SCiNiTO | ||||
| Authors | ||||
Ahmed Aly Nashat   1; Fatma Mazen Aly Mazen 2
| ||||
| 1Fayoum University Faculty of Engineering Electronics and Communication Engineering Dept. | ||||
| 2Fayoum University, Faculty of Engineering, Electronics and Communication Eng. Dept. | ||||
| Abstract | ||||
| Coffee is one of the most consumed beverages in the world and is crucial in the economy of many developing countries. The effectiveness and fast classification of healthy coffee leaves are the most decisive factors in determining their quality for consumers and industrial companies. To separate coffee leaves from the complex real-world background, we proposed an instance segmentation algorithm based on transfer learning and state-of-the-art deep learning algorithm mask regional convolutional neural network (Mask RCNN) for this work. Robusta coffee leaf images dataset (RoCole) is considered in this study. In addition, the VGG Image annotator (VIA) has been used to manually annotate the dataset for coffee leaf segmentation tasks. Resnet101 was adopted as a backbone network, combined with the Feature Pyramid Network (FPN) architecture for feature extraction. The Region Proposal Network (RPN) was trained to create region proposals for each feature map, which used to separate the input image from the complex background. The output image is then fed to a transfer learning-based binary classifier to be classified into one of the two classes. Results reveal that the proposed system has a high-test accuracy of 97.76% for the binary classifier. If the image is classified as unhealthy, it then passes through another segmentation stage based on the HSV color model to highlight the defected areas of the coffee leaf. Instance segmentation results of 148 test images showed that the mean average detection precision rate (mAP@50:95) was 100%, the mean average recall rate (recall@50:95) was 84.5%, and the F1-score was 91.6%. | ||||
| Keywords | ||||
| Convolutional Neural Networks; Instance Segmentation and classification; Coffee Leaf; RoCole; Plant Diseases; VIA | ||||
|
Statistics Article View: 73 PDF Download: 80 |
||||
