DeepLearning_ImageClassification

Flower classification using CNN and data augmentation

Objective

There are three types of flowers in the dataset: Daisy,Rose,Sunflower. The task is to build a CNN model for classification.

Challenges

The dataset size is very small hence the model tends to overfit a lot.

Approach

1. Simple CNN : Overfits a lot due to scarcity of data
2. Simple CNN with data augmentation : Performs well but the losses are a bit erratic
3. Transfer learning with data augmentation : Best model. Performs well with smooth loss curve

Performance

Softmax cross entropy is used. Accuracy is 85%. Loss is 0.40

Project structure:

Folders:

1. data : Contains the dataset with three sub folders
2. dataAugmented : Contains the augmented images. Not used for training purpose. Just to visualize the augmented images to be used.
3. model : The trained model is saved into this folder.
4. templates : GUI related files : index.html, style.css, jqueryScript.js, default.png
5. testImages : Used to store the images along with there predicted class

Files:

1. FlowerClassification.ipynb : Ipython notebook to demonstrate various approaches
2. train.py : The best model obtained is used to train and persist with this file
3. test.py : Flask service which is triggered by the GUI to perform the predictions

Usage:

Create folder data/rose,data/sunflower,data/daisy. Put contents of rose1 and rose2 into rose and sunflower1 and sunflower2 into sunflower.

Training: python train.py 30

30 : epochs

Test: python test.py

Flask service will start on http://127.0.0.1:5000. This is used in the jqueryScript.js to perform the prediction

Open the index.html in templates folder and trigger the service

Reference links:

Bootstrap:

https://bootsnipp.com/snippets/76778

https://www.w3schools.com/bootstrap/bootstrap_templates.asp

Flask:

https://www.tutorialspoint.com/flask

Refer Image Classification Usage.pdf for complete usage guide with screen schots