Karl-Erik/Estonian_Bird_Classifier

Introduction

Hello, and welcome to the Estonian Bird Classifier model page! This model was created by Karl-Erik Kanal as a part of his Bachelor's thesis and can recognise 50 common Estonian bird species.

About the model

The model estonian_birds_classifier is a pretrained InceptionV3 model on ImageNet weights that has been trained using transfer learning to recognise 50 bird species that can be found in Estonia. The complete list of birds that it can classify can be found in the label map provided with the model.

The model was trained and tested using a custom-made dataset for this model, with 5926 images in the training set and 1064 images in the test set. On the test set, the model achieved 74% accuracy, 89% Top-3 accuracy and 91% Top-5 accuracy with a loss of 1.119. In comparison, the Google AIY bird classifier that can recognise 45 species of the 50 achieved 71% accuracy on the test set with the 5 species taken out, while this model achieved 75% accuracy with the same 45 species.

How to use

The model requires the images to be resized to 150 x 150 and normalized before predicting. You can use the ImageDataGenerator class from keras.preprocessing.image to achieve this. The images should be well-cropped to achieve the best results.

You can load in the model like you would load in other Keras type models by using the load_model function. You can load the labels into a dataframe using pandas read_csv. The labels of the birds are provided both in Estonian and in Latin.

model = keras.models.load_model('estonian_birds_classifier.h5')
labels = pd.read_csv('ebc_labelmap.csv', sep=";")