Abstract : Object recognition is a complex problem in the field of computer vision and artificial intelligence with many different solutions including the use of Neural Networks and Deep Learning. This thesis investigates the use of Deep learning for object recognition and improvements to the deep learning methodology using methods such as dropout to reduce overfitting. Improvements to the dropout method are made using selective neuron dropping instead of random neuron dropping used in traditional uniform dropout to improve results up to 1.17% on the CHARS74K dataset using neuron output variance to select neurons for dropping. This thesis also explores the use of Deep Convolutional Neural Networks to create a rotation invariant neural network by using parallel convolutional layers with tied weights with filters presented at different rotations, combining results using a winner takes all pooling method to produce rotation invariance. Results of this method showed great improvement over benchmarks on rotated test data showing a 52.32% increase in accuracy on the MNIST dataset and a 36.44% accuracy increase on the CHARS74K dataset. This work was furthered by attempting to introduce a scaling method into the filter rotation to reduce data loss and blurring by the rotation method, however it was found that results using this method worsened the accuracy of the network compared to benchmarks not using the method, showing a decrease in accuracy of 0.17% on the MNIST dataset and a decrease of 3.68% on the CHARS74K dataset.
© Erik Barrow