Application of several techniques of data collection and labeling, Machine Learning and Deep Learning to solve the problem, which aims to classify an image in one of three classes: day (the first two images), transition (the third and fourth ) and night (last two). The accuracy obtained in the best model for this problem was 98%.
For more information, read Report.pdf (in Portuguese) or contact me.
- Transient (http://transattr.cs.brown.edu/);
- AMOS (http://cs.uky.edu/~jacobs/datasets/amos/);
- Barcelona (http://www.cs.unc.edu/~jtighe/Papers/ECCV10/);
- Places (https://ieeexplore.ieee.org/document/7968387).
(I) - Histogram, (II) o GLCM. F1-score over 20% of the data - cross validation.
| Classifier | I | II | I + II |
|---|---|---|---|
| Linear SVM | 84.81% | 88.70% | 90.85% |
| Decision Tree | 82.17% | 82.52% | 85.39% |
| Random Forest | 88.89% | 88.87% | 90.81% |
| Neural Network | 87.72% | 92.18% | 91.74% |
| Naive Bayes | 79.05% | 68.89% | 69.66% |
| Logistic Reg | 84.80% | 86.66% | 90.69% |
(III) - Local Binary Pattern (LBP)
| Classifier | III | II + III | I + II + III |
|---|---|---|---|
| Linear SVM | 82.82% | 89.39% | 90.15% |
| Decision Tree | 76.66% | 84.44% | 84.67% |
| Random Forest | 80.00% | 91.36% | 90.72% |
| Neural Network | 84.92% | 92.89% | 92.28% |
| Logistic Reg | 83.18% | 87.98% | 88.57% |
| Classifier | ResNet50 | IncV3 | Xcp |
|---|---|---|---|
| Linear SVM | 94.77% | 92.66% | 91.85% |
| Random Forest | 91.87% | 79.39% | 76.17% |
| Neural Network | 95.73% | 91.68% | 90.20% |
| Logistic Reg | 95.24% | 68.79% | 80.07% |
Results for each classifier:
| Classifier | accuracy | f1-score |
|---|---|---|
| 1 | 94.39% | 94.39% |
| 2 | 95.79% | 95.76% |
| 3 | 94.20% | 94.20% |
| 4 | 95.98% | 95.96% |
| 5 | 94.84% | 94.80% |
| 6 | 97.51% | 97.50% |
| 7 | 97.45% | 97.44% |
Meta classifier:
| Classifier | accuracy | f1-score |
|---|---|---|
| Linear SVM | 97.13% | 97.13% |
| Decision Tree | 98.08% | 98.08% |
| Random Forest | 98.02% | 98.02% |
| Neural Network | 96.68% | 96.70% |
| Logistic Reg | 96.81% | 96.82% |
| Train | Val |
|---|---|
| Loss | Accuracy |
| 0.2789 | 90% |
| Train | Val |
|---|---|
| Loss | Accuracy |
| 0.3222 | 97.14% |
| Predict Day | Predict Night | Predict Trans | |
|---|---|---|---|
| Day | 589 | 0 | 1 |
| Night | 3 | 562 | 15 |
| Trans | 9 | 7 | 384 |
-
PARAMETERS
- In order to correctly execute the codes, some parameters must be configured as described below.
- In util.py
- Paths of the datasets;
- num_cores: used in processing. All the 'hot' functions of this work were paralleled, and all use this parameter;
- In train.py
- MODEL_ACTION
- TEST_ALL_CLASSIFIERS - Tests all classifiers under selected features;
- TEST_BIG_CLASSIFIER - Train/Test the 'big' classifier presented and explained in the paper;
- IMAGE_ROTULATION - Labeling of New images with the 'big' classifier, being the image added to the dataset only if all classifiers labeled with the same class;
- AMOS - Labeling data of the dataset AMOS;
- MODEL_ACTION
-
The files have documentation summarized in all their functions.
-
Files
- deep_learning_train.py: Train CNNs with fine tuning;
- feature_extraction_CNN.py: Feature extraction with CNNs;
- feature_extraction.py: Feature extraction with histogram, co-matrix e LBP;
- process_dataset_AMOS.py: Data processing functions for the AMOS dataset;
- process_dataset_nexet.py: Data processing functions for the Nexet dataset;
- process_dataset_transient.py: Data processing functions for the Transient dataset;
- train.py: Train all classifiers (the features have already been extracted, except in the case of deep models);
- util.py: Several "general purpose" functions that were required during this work;
- vgg16_places_365.py: Pre-trained model vgg16_places_365.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent