Javascript must be enabled for the correct page display

The effect of a stacked MLP trained on output probabilities in a patch-based image classification system

Guchte, L.A. van de (2016) The effect of a stacked MLP trained on output probabilities in a patch-based image classification system. Bachelor's Thesis, Artificial Intelligence.

[img]
Preview
Text
AI_BA_2016_LennartVanDeGuchte.pdf - Published Version

Download (716kB) | Preview
[img] Text
Toestemming.pdf - Other
Restricted to Backend only

Download (77kB)

Abstract

This thesis presents a patch-based image classification system in combination with anensemble technique named stacking. In the standard model an image is reshaped into a collectionof patches, from these patches feature vectors are extracted using the Histogram of OrientedGradients (HOG). Then a simple 3-layered multilayer perceptron (MLP) is used for classification.Classification is done by applying the sum voting technique, here the output probabilities of eachpatch in an image are summed up. The image is then assigned to the class with the highest sumof probabilities. During the training of this MLP dropout is applied to prevent the networkfrom overfitting. In the ensemble model an extra MLP is stacked after the usual MLP. Here themaximum, average or minimum output probabilities from all patches in an image are computedand serve as input values for the stacked MLP. Additionally, we varied with a pooling methodthat splits an image into quadrants. The patches of each quadrant are then used to train differentMLPs, the output probabilities of these MLPs serve as input values for the stacked MLP. Tomeasure the accuracy of these models the two widely known image recognition benchmarksMNIST (handwritten digits) and CIFAR-10 (objects) are used for experimentation. The standardmodel shows good performances on the MNIST dataset (99.45% correctly classified images). Theaddition of stacking and/or pooling results in a decrease in accuracy of 0.20-1.63%. On CIFAR-10the standard model also outperformed all other models (68.30%) with a decrease of 1.17-12.97%when stacking and/or pooling is used.

Item Type: Thesis (Bachelor's Thesis)
Degree programme: Artificial Intelligence
Thesis type: Bachelor's Thesis
Language: English
Date Deposited: 15 Feb 2018 08:25
Last Modified: 15 Feb 2018 08:25
URI: http://fse.studenttheses.ub.rug.nl/id/eprint/14698

Actions (login required)

View Item View Item