| Titre : | A CNN based architecture for multispectral image classification: Application on Dates Sorting |
| Auteurs : | Sarah Setta, Auteur ; Abdelhamid Djeffal, Directeur de thèse |
| Type de document : | Monographie imprimée |
| Editeur : | Biskra [Algérie] : Faculté des Sciences Exactes et des Sciences de la Nature et de la Vie, Université Mohamed Khider, 2022 |
| Format : | 1 vol. (80 p.) / couv. ill. en coul / 30 cm |
| Langues: | Anglais |
| Mots-clés: | Convolutional Neural Networks (CNN), Deep Learning, Keras, TensorFlow,Thermal image, quality control |
| Résumé : |
Dates are a popular and abundant fruit in the Middle East and North Africa, with a growing international presence. There are numerous types of dates, each with unique characteristics. Sorting dates is an important process in the date industry, but it can be a time-consuming task when done manually. The rapid advancement of AI technologies and their applications in agriculture opens up completely new avenues for intelligent systems to better forecast trends and assist farmers in making decisions. A variety of advanced computational techniques are used to recognize and identify the types and quality of various fruits using a variety of methods. Machine vision is used for grading and sorting in multispectral imaging, monochrome imaging, and color imaging. Deep Neural Networks (DNN) are extremely effective at identifying and classifying fruit images. They outperform other machine learning algorithms in terms of accuracy. Convolutional Neural Networks are the most common type of Artificial Neural Network and are considered to be the most efficient Deep Learning Algorithms. Convolutional Neural Networks (CNNs) are intended to map a single image data input. The majority of the literature focuses on improving the architecture itself, feeding only one fruit side as input, which may result in overfitting. Others employ multi-input CNNs on all fruit sides. However, given the complexity of CNN, this method is time-consuming because each side passes through the CNN. As a result, in such cases, image processing is the gold standard for overcoming this issue. but it must still be chosen as the best processing tool.As a solution, we proposed a new simple yet effective processing method thatexplores all fruit faces, based on thermal image and other characteristics as the date weight, combines them as one image input using image processing steps, and applies two different CNN models to the final data result. |
| Sommaire : |
Contents Acknowledgements ii Dedication iii Abstract iv Resume v Contents vii List of Figures xi List of Tables . xiv General Introduction 1 1 Insights into deep learning 5 1.1 Introduction . 5 1.2 Machine learning 6 1.2.1 Types of Machine learning algorithms 6 1.3 Deep Learning 9 1.3.1 Data organization in Deep Learning algorithms10 1.3.2 Deep learning applications 11 1.4 Deep learning algorithms: 12 1.4.1 Artificial Neural Networks (ANNs) 12 1.4.2 Deep Neural Networks: Types and architectures17 1.4.3 Colvolutional neural network (CNN) 20 1.5 Conclusion . 26 vii2 Dates Fruits Quality Control: Methods Overview and Review 28 2.1 Introduction 28 2.2 Quality Control 28 2.2.1 The importance of quality assurance29 2.2.2 Benefits of quality control [1]29 2.2.3 Characteristics of a control [40]30 2.2.4 Quality types 30 2.2.5 Relevance of control by an artificial vision: 31 2.3 Deployment [32] . 31 2.3.1 Machine Vision . 31 2.3.2 Image recognition 31 2.3.3 Image acquisition 32 2.3.4 Image processing 32 2.3.5 Images classification 33 2.4 The image 34 2.4.1 Definition . 34 2.4.2 Features of a digital image 34 2.4.3 Types of images 37 2.5 Feature extraction 38 2.5.1 Local features 38 2.5.2 Global features38 2.6 Thermal imaging (TI) as a technique for process analysis in the food industry [40] 39 2.7 Related works 39 2.8 Conclusion 40 3 Conception of the architecture for the classification of dates fruits 43 3.1 Introduction 43 3.2 Problem statement43 3.3 General architecture44 viii3.4 Detailed architecture 45 3.4.1 Step 01: data acquisition [40] 47 3.4.2 Step 02: data processing 49 3.5 Database organization 51 3.5.1 Data split 51 3.5.2 Stratified k fold cross-validation 51 3.6 Model conception . 52 3.6.1 Inception-based transfer learning[31]52 3.6.2 Deep learning Model [37] 52 3.7 Performance evaluation [22] 53 3.8 Conclusion . 54 4 Experimental Study 56 4.1 Introduction 56 4.2 Development tools and programming languages 56 4.2.1 Python .. 56 4.2.2 PyCharm 57 4.2.3 Flask 57 4.2.4 Google Colab 58 4.2.5 TensorFlow 58 4.2.6 Keras . 59 4.2.7 Numpy 60 4.2.8 Flask-dropzone 60 4.3 Database Description 61 4.4 Application 63 4.4.1 Data Preprocessing 63 4.4.2 Input data 67 4.4.3 Design the CNN to classify Dates68 4.4.4 Results Discussion and comparison 76 4.5 Model deployment, processing visualization . 77 ix4.6 Conclusion 79 General Conclusion 80 |
Disponibilité (1)
| Cote | Support | Localisation | Statut |
|---|---|---|---|
| MINF/706 | Mémoire master | bibliothèque sciences exactes | Consultable |
