This project is based on WPI RBE 549 Computer Vision.
The aim is to create an algorithm to distinguish images of dogs from cats. During the project two classifiers are used, CNN (Tensorflow) and SVM (Scikit-learn).
Dataset: Kaggle’s official competition dataset Asirra;
In this project, our team also use several enhancement methods together with Support Vector Machine to reduce the overfit- ting problem. Firstly, instead of using the pixel value as sample feature, extract HOG features from images. Secondly, apply two ensemble learning method, Adaboost and Bagging, to increase the accuracy rate of SVM classifier.
And as for CNN, we learn from CIFAR-10 as it has a tutorial for building a CNN network using CIFAR-10 to classify RGB 32*32 pixel images across 10 categories. And the advantage of this model is that it is complex enough to exercise much of TensorFlows ability to scale to large models, while at the same time, its small enough for our personal computer to train in limited time, which is really ideal for experimenting new techniques and ideas.
-
- SVM Result
-
- CNN Net Used
