This repository contains the scripts used in retraining the Inception v3 model to classify images of human faces into five basic shapes: heart, oblong, oval, round, and square. The repository also contains the scripts used to benchmark it to traditional classifiers using features derived from facial landmark coordinates generated using OpenCV and DLIB pre-trained models.

Training images used are downloaded via Google image search and may have copyright constraints; I can give you a copy of the images, bottleneck files, and/or feature files by request if you promise not to redistribute and to only use for academic purposes. You can send me an e-mail at [email protected] if you are interested.

CLASSIFY_FACE.PY
This script runs the re-trained Inception model to classify a single or a batch of images

CLASSIFY_FACE_CONFUSION.PY
Similar to classify_face.py but generates a text file of results and a confusion matrix

EXTRACT_FEATURES.PY
This script detects the face(s) in the image, specifies the bounding box, detects the facial landmarks, and extracts the features for training

PROCESS_IMAGE.PY
This script contains a couple of image pre-processing and augmentation functions like squaring an image, filters, blurs, zoom, rotate, flip, and recolor, etc

RETRAIN_CMDGEN.PY
This script generates the Windows CMD command to re-train the Inception v3 model that tees CMD line prompts into a text file; Set up the needed files and directories then run in the CMD line to retrain the model.

RETRAIN_v2.PY
#Slight modifications in defining the overall test set to include all images, resolved issue of "doubling-counting" of validation and test images, added CMD line arguments on where to save useful info as txt file

TRAIN_CLASSIFIERS.PY
This script trains the LDA, SVM-LIN, SVM-RBF, MLP, and KNN classifiers for a set of training set sizes

PAPER.PDF
A short paper describing the methodology and experimental results

bottlenecks.rar
COntains the bottleneck files (TXT) of the 500 images in the dataset. Note: Bottleneck files are the vector representation of the images at before the last layer of the Inception model.

features.txt
COntains the feature vectors of the 500 images in the dataset used in the LDA, SVM, KNN, and MLP classifiers.