Javascript must be enabled for the correct page display

Weighting Patches of a Multilayer Perceptron for Image Classification

Knigge, L.S. (2017) Weighting Patches of a Multilayer Perceptron for Image Classification. Bachelor's Thesis, Artificial Intelligence.

AI_BA_2017_Knigge.pdf - Published Version

Download (776kB) | Preview
[img] Text
Toestemming.pdf - Other
Restricted to Backend only

Download (80kB)


Nowadays, image classification becomes increasingly indispensable in our modern day lives, with a wide range of applications such as in identifying handwritten postal codes or object recognition in computer vision. This thesis describes a method to improve image classification with patch-based multilayer perceptrons (MLP). A weight is applied to each patch that is produced by a second MLP. This MLP is also trained on patches, but it will learn the weight, which is the probability the MLP awarded to the correct class. For the test images the patches are extracted and fed into both MLPs which results in output vectors for the first, and weights for the second MLP. The weights are then applied to the output vectors which are then combined, using sum voting, to produce a classification. The purpose of weighting the patches is to improve the accuracy of the system. To test this, the MNIST (handwritten digits) and the CIFAR-10 (small images) datasets are used. We compare the accuracy rates with and without weighting as well as the certainties for these classifications. This study demonstrates that even with a simple MLP with one hidden layer, error rates lower than 1% on the MNIST dataset can be achieved. However it does not surpass the state-of-the-art techniques. The results also display that the new algorithm does not improve the image classification accuracy of the system on both datasets.

Item Type: Thesis (Bachelor's Thesis)
Degree programme: Artificial Intelligence
Thesis type: Bachelor's Thesis
Language: English
Date Deposited: 15 Feb 2018 08:30
Last Modified: 15 Feb 2018 08:30

Actions (login required)

View Item View Item