Image Classification

The objective of this project is to build a Image classifier that identifies the correct orientation of the image.

About the dataset The dataset consists of micro-thumbnail images (8x8 pixels) represented as 192 dimensional feature vectors (8x8x3, with each pixel represented as a separate feature for Red, green and blue color densities). The training datafile has one row per image, with each row formatted as: photo_id correct_orientation r11 g11 b11 r12 g12 b12 ... The testset has similar format except for the first 2 columns

About the code

The orient.py module has the main program that accepts arguments for path to the training dataset and test datasets, algorithm (nnet)

The program can be run as below:

./orient.py train-data.txt test-data.txt nnet This invokes the Neural Network algorithm implementation in nnet.py

The neural network has one hidden layer. Implemented Cross-validation based model selection to train the model with different settings for the hidden_layer_nodes parameter Best parameter is then used and the model trained again on the full dataset, that can be used to predict on unseen datasets.

A brief Report

Getting accuracy on an average about 70-75% for different configurations of the nodes in the hiddlen layer. Observing very small improvements in accuracy for NNs with higher number of nodes but seems to mostly plateau for anything with 16 and above nodes in the hiddel layer The training is also pretty fast as well as the classification is fast with NNs. Seems to be a good algorithm for predicting with the given dataset, as the accuracy and computation speed seems to be pretty good