Bird-Species-Classifier

Deep learning model to classify input of image of a bird to the respective species., and return top 3 predictions for the input image

Introduction

The model is trained on a dataset which has 250 different classes, hence can classify into 250 different species. On running main.py, we can input an image of a bird, and the model will classify it into it's respective species.

Dataset

The dataset is derived from Kaggle, and is linked below.

• It is structured to have training, test and validation folders.
• It consists of 250 distinct classes, with images in all three folders.
• It also includes a consolidation folder, which consists of all images from all the sub-folders.

Dataset Link: https://www.kaggle.com/gpiosenka/100-bird-species

Data has been scaled down to 224 x 224, in order to have uniformity and ease while training the model.

Model Features

Resnet 152

• ResNet152V2 is the pretrained model that is used for this particular classification.
• Model used is trained with the weights of imagenet dataset.
• Using 3 color channels 'rgb'

Artificial Deep Neural Network

• Consists of GlobalAveragePooling layer to reduce dimensionality and to flatten the output data from feature extraction layers.
• Usage of Dropout to prevent overfitting and improving accuracy.
• Usage of BatchNormalization to stabilize learning process and reduce number of epochs required for higher accuracy.
• Output layer has 250 neurons, for classification into 250 classes.
• softmax is used on the last layer to return it as probability density, for easier classification.
• ReLu is used as it is effiient and easy to train.

Compiling and Training Model

• Model is compiled with loss function as categorical_crossentropy
• Optimizer used is adam. learning rate is modified to 1e-04.
• Model is trained for 50 epochs for a decent accuracy.

Accuracy and Overfitting

• Dropout layer is used generously to assure high performance
• BatchNormalization to reduce overfitting
• Data Augmentation is used to increase size of dataset, to obtain a more comprehensive data, to us layer.

Accuracy Results

At the last epoch:

Training Accuracy	Validation Accuracy
88.76%	94.48%

Graphs

• Usage of matplotliob.pyplot
• Graph that plots Accuracy vs Epochs

Accuracy Graph:

• Graph that plots Loss vs Epochs

Loss Graph:

• As can be made out from the graph, the model is not overfitting, and has a great accuracy.

Requirements

Libraries required are:

tensorflow
numpy
PIL

Usage

• Run main.py from the command line.
• Input the path of the image when prompted.
• The output will display the input image as well as the top 3 predictions with their respective probabilities.

Sample Execution

The model accurately predicts with high confidence(99.97%) that the input image is a bald eagle, and the other two top probabilities are negligible.