ChessPiecePrediction

Implemented Machine Learning and Convolutional Neural Networks to predict chess pieces.

Overview

In this project we used the chessman image dataset from kaggle.com. The dataset can be found here: Dataset
We wanted to use this data set to train and test different AI models and Techniques that result in good prediction accuracy, and were able to be used to predict future images and classify them into various chess piece categories.

CNN

As one approach to classifying chess images, we used Convolutional Neural Networks (CNN) with transfer learning. One of the major problems with image classification is the large number of features that come with it. Each pixel of an image represents a feature. For a small 600x600 pixel RGB image, there are 1,080,000 features. This makes feeding such features to neural networks impractical. CNNs are useful in extracting the useful parts of the image, effectively shrinking down the number of features to be fed into the neural network. We can also shrink down the size of the image before processing with CNNs.

My Contribution

PCA

We implemented principal component analysis. Due to the high dimensionality nature of our data, we needed to try a dimension reduction approach and combine it with other prediction models. Based on our research of multiple Dimensionality Reduction Techniques Python & PCA guide we decided to use Principal Component Analysis (PCA) to reduce dimensionality of our data.

we also used a built in function to normalize the data, which is usually recommended/required when running a PCA analysis: we used the normalize implementation within sklearn.preprocessing library In addition, we did our dimensionality reduction using both grayscale and original colors. Preprocessing-images-dim-redux was used as a reference point when devising our approach. PCA was fit on the training data set which was predetermined using a 80:20 train/test ratio. We then attempted to find the principal components that captured 95% variance of the dataset. We then transformed our training feature dataset (train_X) and testing X feature dataset (test_X) using the determined components.

Example

Machine Learning

Project Takeaways

Based on the approaches we explore in this project, we think CNN are better at accurately classifying chess pieces compared to using a Principle Component Analysis approach. Other things we wanted to consider for the future to further test and/or improve performance would be:

• Expand dataset to include non chess images
• Train CNN model on images both with and without static noises
• Try other dimensionality techniques other than PCA. Some ideas would Non-negative Matrix Factorization and Chi^2 dimension reductionality methods
• In case of PCA approach, resizing the image shapes in a different way to feed PCA and machine learning methods. Currently converting images into numpy array which had initial shape of (samples, width, height, RGB) to (samples, width x height x RGB) array.
• Use a gridsearch to tune model hyper-parameters more effectively for our dataset.
• Experiment with XGBoost models
• Overall, the PCA approach didn’t result in favourable outcomes, but has room to be improved. Nevertheless, it is proven to be better than random prediction since for our case random would ~ 16.67%.