m0rp43us / aerial-cactus-identification

Cactus Identification: A Walkthrough from CNN to Transfer Learning with Pre-Trained Model VGG16.

Cactus Identification: A Walkthrough from CNN to Transfer Learning with Pre-Trained Model VGG16

This project was my introduction to the world of Computer Vision using Deep Learning. In this write-up, I will do a walkthrough of the whole classifier making process, including:

• Data preprocessing

• Build a Convolutional Neural Network

• Evaluate the model

• Make tweaks to improve performance: data augmentation, progressive resizing

• Apply transfer learning with a pre-trained model (VGG16)

• Test the model by making prediction on first 10 images from the test set

• Experiment with interactive testing using OpenCV

Full codes can be found here. NBViewer Link.

Data Description

This dataset contains a large number of 32 x 32 thumbnail images containing aerial photos of a columnar cactus (Neobuxbaumia tetetzo). The file name of an image corresponds to its id.

I will attempt to create a classifier capable of predicting whether an images contains a cactus.

Data Overview

Visualize Images

Let's take a look at a few sample images from the train set:

Data Preprocessing

In this section, we create data generators for both train and test dataset. We rescale pixel values from [0, 255] to [0, 1], then perform data augmentation to all images. Full Python script for the data preprocessing process can be found here.

Model Building

Baseline Convolutional Neural Network Architecture

Model Evaluation: Loss/Accuracy Graphs

Transfer Learning with Pre-Trained Model VGG16

In this section, we load the pre-trained model, freeze all layers except the top FC layers, then we add the top layers accordingly. Let's take a look at the model architecture after applying transfer learning:

Model Evaluation: Loss/Accuracy Graphs

Test Our Model

In this section, we write a Python script to test our model by making prediction on the first 10 images of the test set. Full Python script can be found here.

Conclusion

In this project, I did a walkthrough of the whole classifier making process, including data preprocessing, building a Convolutional Neural Network, evaluating the model, making tweaks to improve performance: data augmentation, progressive resizing, applying transfer learning with a pre-trained model (VGG16), testing the model by making prediction on first 10 images from the test set, and experimenting with interactive testing using OpenCV.

Computer Vision is a wildly developing field. Since the conclusion of this project, I have been focusing heavily on researching recent advances in convolutional neural network architectures, and hopefully I can apply what I've learned in more interesting projects like this in the future.