Identification and Classification of Seed Coat Sculptures using Deep Learning

다국어 초록 (Multilingual Abstract)

In this study, we worked on the rarest biological data analysis. This research aimed to compare between the human-based classification and machine-based classification. Most common biological data analysis of the work is the identification of plant diseases. The main contribution of this work was to study the suitable use of human-based methods by applying machine-based methods to automatic classification of Allium seed walls (anticlinal and periclinal). Thus, we briefly explain two methods of work.
Seed coat characteristics provide important information in the taxonomy of Allium. The seed coat surface was more diverse. In this work, we explored new classification for Allium seed coats, using five different networks CNN, AlexNet, GoogleNet, ResNet50, and VGG16 on Scanning Electron Microscopy (SEM) images over 100 species of Allium seeds.
We considered seven common types of anticlinal wall, classified as irregular curved, irregular curved to nearly straight, straight, S, U, UO, and omega. Also, five custom types of periclinal wall distinguished at granule, small verruca, large verruca, marginal verruca, verrucate verrucae. GoogleNet achieved the highest classification accuracy, with 90.4% for the anticlinal wall, and VGG16 achieved the highest accuracy, with 98.1% for the periclinal wall. Therefore, we implemented model performance in more than three, four, and five methods. As a result of the study, more than three methods are more benefit able from more than four and five methods. The implemented models achieved an anticlinal wall classification suitable rate in GoogleNet architecture, and the periclinal wall classification high accuracy rate is VGG16 architecture. In conclusion, using our own machine-based method, we studied the seed coat of species of Allium on our own samples to see if the results of the machine-based method with the human-based match.
Seed coat sculptures, including anticlinal and periclinal walls, are great taxonomic importance. In this study, we identified seed coat patterns of Allium seeds using five deep learning methods namely, CNN, AlexNet, GoogleNet, ResNet50, and VGG16 for the first time. Selected images of seed coat patterns from over 100 Allium species reported in previously published literature and data from our samples were classified into seven types of anticlinal (irregular curved, irregular curved to nearly straight, straight, S, U, UO, and omega) and five types of periclinal walls (granule, small verruca, large verruca, marginal verruca, and verrucate verruca). The results revealed that GoogleNet and VGG16 achieved the highest classification accuracy of 90.4% for the anticlinal wall, and VGG16 achieved the highest classification accuracy of 98.1% for the periclinal wall. Moreover, more than three, four, and five methods were combined, and their performance was investigated. Combining more than three methods was the most advantageous. The models achieved a suitable anticlinal wall classification using GoogleNet and periclinal wall classification using VGG16. In conclusion, using the machine-based method, we studied the seed coat of species of Allium on our own samples to see if the results of the machine-based method with the human-based classification match.

목차 (Table of Contents)

• 1. Introduction .............................................................................................. 1
• 1.1. Genus Allium L. ................................................................................... 1
• 1.2. Seed testa sculpture .............................................................................. 1
• 1.3. Macro- and micro morphological characters ....................................... 4
• 1.4. Scope of thesis ...................................................................................... 6
• 2. Background study ..................................................................................... 8
• 2.1. Machine Learning and Deep Learning basics ..................................... 8
• 2.1.1. Machine Learning .................................................................... 8
• 2.1.2. Deep Learning .......................................................................... 9
• 2.2. Convolutional Neural Network ......................................................... 10
• 2.2.1. Input representation ............................................................... 11
• 2.2.2. Edge detection ........................................................................ 12
• 2.2.3. Stride and Padding ................................................................. 12
• 2.2.4. Convolution Layer ................................................................. 14
• 2.2.5. Rectified Linear Units ............................................................ 15
• 2.2.6. Pooling Layer ......................................................................... 15
• 2.2.6.1. Max pooling ……………………………………….... 16
• 2.2.6.2. Global average pooling …………….………….…….. 16
• 2.2.7. Fully Connected Layer ........................................................... 17
• 2.2.8. Training a network ................................................................. 17
• 2.2.8.1. Loss function……….…………………………...…… 18
• 2.2.8.2. Gradient descent ……………………………………... 18
• 2.3. CNN architectures ............................................................................ 18
• 2.3.1. AlexNet .................................................................................. 18
• 2.3.2. GoogleNet .............................................................................. 19
• 2.3.3. ResNet .................................................................................... 20
• 2.3.4. VGG16 ................................................................................... 20
• 3. Identification of Seed Coat Sculptures using Deep Learning ............. 22
• 3.1. Datasets.............................................................................................. 22
• 3.2. Data pre-processing ........................................................................... 24
• 3.2.1. Anticlinal wall detection ........................................................ 24
• 3.2.2. Periclinal wall detection ......................................................... 26
• 3.3. Data augmentation ............................................................................. 27
• 3.4. Creating training and test sets ............................................................ 28
• 3.5. DL models ......................................................................................... 28
• 3.6. Combined models .............................................................................. 30
• 3.7. Results ............................................................................................... 31
• 3.7.1. Comparison of architecture performance ............................... 31
• 3.7.2. Identification of the anticlinal wall using each model ........... 32
• 3.7.3. Identification of the periclinal wall using each model ........... 34
• 3.7.4. Results of the combined models ............................................ 35
• 3.7.5. DL for detecting the anticlinal walls ...................................... 38
• 3.7.6. DL for detecting the periclinal walls ..................................... 39
• 3.7.7. Combined model results ........................................................ 40
• 3.8. Discussion .......................................................................................... 42
• 3.9. Summary............................................................................................ 43
• 4. A multi-considered seed coat pattern classification of Allium L. using unsupervised machine learning ....................................................... 45
• 4.1. Proposed method ............................................................................... 45
• 4.2. General module ................................................................................. 45
• 4.2.1. Preprocessing module ............................................................ 45
• 4.2.2. Augmentation module ............................................................ 46
• 4.2.3. Shuffle raw data ..................................................................... 46
• 4.2.4. Calculate the number of clusters ............................................ 47
• 4.2.5. Clustering module .................................................................. 48
• 4.3. Performance analysis ......................................................................... 48
• 4.3.1.Visualization module .............................................................. 50
• 4.4. Result comparison ............................................................................. 50
• 4.4.1. Comparison between human-based and machine-based results ..……………………..……………………………………………. 51
• 4.4.2. Proposed clustering ................................................................ 54
• 4.5. Discussion .......................................................................................... 58
• 4.6. Summary............................................................................................ 59
• 5. Conclusion and Future work ................................................................. 60
• 5.1. Conclusion ........................................................................................ 60
• 5.2. Future work ....................................................................................... 61
• REFERENCES ........................................................................................... 65
• ABSTRACT ................................................................................................ 73