More than 45 million people watch birds around their homes and away from home, according to the findings of the U.S. Fish & Wildlife service. Nowadays, bird species identification is seen as a mystifying problem which often leads to discombobulation and uncertainty. Many people visit bird sanctuaries to look at the birds, while they barely recognize the differences between various species of birds and their characteristics. Understanding such discrepancies between species can increase our knowledge of birds, their ecosystems and their biodiversity.
Bird watching is the practice of observing birds in their natural environment. The objective of this project is to assist birdwatchers as well as ornithologists correctly identify different species of birds with ease.
To achieve this, we implement Supervised Machine Learning techniques such as deep learning and classification algorithms to develop a model to accurately identify 225 different bird species given an image of a bird. In this project, we utilize various ImageNet Pre-trained Neural Network models for classification of 225 Bird species. Additionally, we also use Machine Learning models such as Logistic Regression and SVM to classify the images using the features extracted using various techniques.
For feature extraction, we implemented the following techniques:
1. HU Moments
2. Haralik Textures
3. Color Historgram
4. VGG16 for feature extraction
5. Bag of Visual Words
The dataset consists of over 33 thousand 224x224 color images of 225 different bird species.
For a detailed study, kindly refer the document: "DS5220_Project_Report.pdf"
The following are the code files:
1. Log_Reg_Flattened.ipynb - Logistic Regression using flattened feature vectors
2. Log_Reg_Reduced.ipynb - Logistic Regression using only 50 images per class VGG16 extracted features
3. Saving_Extracted_Features.ipynb - Saving the VGG16 extracted features to files
4. SVM_Flattened.ipynb - Attempt to train SVM on flattened features - did not execute even after 40 hours
5. SVM_GRIDSEARCH_Reduced.ipynb - Hyperparameter tuning for reduced data - only 50 images per class
6. SVM_LogReg_Load_Saved_Models.ipynb - Loading the best performing saved SVM and Logistic Regression models and displaying accuracy
7. SVM_training.ipynb - Training the SVM model with complete VGG 16 extracted features
8. pretrained_cnn_models - Contains code to train various ImageNet pre-trained models offered by PyTorch. They include variations of densenets, squeezenets, VGG, resnets etc.