Classification of Dog

Contents

Objective: Building a model to classify between 133 different breeds of dogs and identify them

A pipeline is built to process real-world, user-supplied images. Given an image of a dog, the algorithm will identify an estimate of the canine's breed. If supplied an image of a human, the code will identify the resembling dog breed. The test accuracy target for the CNN is 90% i.e., the model identifies the dog breed 9 times out of 10 correctly. The accuracy metric on the testing dataset is used to measure the performance of our models

Step -1: Import required libraries

• Utility libraries - random (for random seeding), timeit (to calculate execution time),os, pathlib, glob(for folder and path operations), tqdm (for execution progress), sklearn (for loading datasets), requests and BytesIO (load files from the web)
• Image processing - OpenCV (cv2), PIL
• Keras and fastai for creating CNN
• matplotlib for viewing plots/images and numpy for tensor processing

Step 0: Import Datasets - The datasets were provided by Udacity. 

• Dog Images - The dog images provided are available in the repository within the Images directory further organized into train, valid and test subfolders
• Human Faces - An exhaustive dataset of faces of celebrities have also been added to the repository in the lfw folder
• Haarcascades - The algorithm uses the Haar frontal face to detect humans. So the expectation is that an image with the frontal features clearly defined is required
• Test Images to check algorithm - A folder with certain test images have been added to be able to check the effectiveness of the algorithm
• Pre-computed features for networks currently available in Keras (i.e. VGG19, InceptionV3 and Xception) will be made available from S3
• The folders in the repository have been organized as lfw (containing human images), images (containing dog images organized into train, valid and test sub folders), Haarcascades (containing Haarcascade files), test_images (containing 8 images to check the algorithm).
• The files are Readme.md, dog_breed_classifier.ipynb (the main iPython notebook), extract_bottleneck_features.py (a file to extract the predictions from the keras transfer learning models), and sample_cnn.png (an illustrative CNN model)

Step 1: Detect Humans

• The face_detector function takes a string-valued file path to an image as input and returns True or False depending on whether a human face is detected in an image or not

Step 2: Detect Dogs

• The dog_detector function, returns True if a dog is detected in an image (and False if not)

Step 3: Create a CNN to classify Dog Breeds (from scratch)

• A CNN model with 4 convolutional layers alternating with max-pooling layers, dropout and batch normalization with Keras has been fit for 10 epochs for a test accuracy of 6.7%.  

Step 4: Use a CNN to classify Dog Breeds (using Transfer Learning)

• Bottleneck features of VGG16 was used to generate a transfer learning model which generated a test accuracy of 48%.

Step 5: Create a CNN to classify Dog Breeds (using Transfer Learning)

• Other Keras models such as VGG-19, ResNet50, InceptionV3 and Xception have also been used for Transfer Learning based models. These models brought the accuracy upto 80+%.
• fastai was also used to create a CNN model. This resulted in test accuracy upto 89%.

Step 6: Write an algorithm

• Given higher accuracy generated by the fastai model, this model was chosen to generate the final predictions
• The predict_breed function takes an input of a file_path and outputs the breed of the dog
• The algo function determines if the provided file_path contains a dog or human or neither
• The provide_output outputs a message based on the predicted species and dog breed

Step 7: Test algorithm

• 6/6 dogs were correctly identified as dogs. 5/6 were accurate breeds. Only 1 dog (a Rajapalayam, a native breed was identified as a Great Dane, possibly because Rajapalayam is not one of the 133 breeds in the ImageNet dataset.
• 2/2 humans were correctly identified as humans and a dog breed was predicted for them

Summary of results

• The final model obtained 89.8% testing accuracy close to the targeted 90%.
• There are a few breeds that are virtually identical and are sub-breeds.
• There's also a possibility of some images being either blurred or having too much noise.
• There's also a possibility of enhancing the quality by additional image manipulation.

A simple web application in Flask could be built to leverage the model to predict breeds through user-input images.

Acknowledgements

StackOverflow, various Kaggle kernels