Project Proposal, CS 174 Spring 2012

Introduction

In the machine learning area, a great deal of research on the image recognition and classification has been performed. Many algorithms and methods of the problems have been developed and well-studied. While having been implemented to human beings a lot, they have not been done much to animals.

Dogs have coexisted with human beings for a long time. The AKC(American Kennel Club) recognizes over 170 individual dog breeds and it categorizes as seven groups according to their historical functionality [1].

In the project, we are going to apply image recognition and classification algorithms to dogs to determine dog breeds from dog pictures. And we are going to compare our performance with the result from Khosla et al. [2].

Methods

We are going to recognize dogs from every image that contains a dog. Fortunately, Stanford data set provides bounding box annotations for all images. However, the dataset is meant to be challenging by containing dogs of different ages/appearance within the same category. Thus, our big picture is that we will implement (or import known) Haar-like feature filters to classify dog images (Weak classifiers). These filters will be tested with AdaBoost learning method. After all steps, The our filters will be combined linearly to obtain the best result (Strong classifier). We might calculate integral image to reduce the calculation cost (but have not decided yet). If we have enough time, We will try to let our program(machine) recognize not only dog but also its breed.

Data set

We will use Stanford Dogs Dataset which contains images of 120 breeds of dogs around the world [2]. It contains dog's images as well as other objects which are needed to be pre-processed to experiment properly.

Schedule

• 4/24 : Make detection wavelet template and window size to create weak classifier
• 4/30 : Construct,(or import) several weak classifiers
• 5/8 : Start testing created weak classifiers with AdaBoost learning method.let our machine learn Strong classifier and milestone write-up
• 5/20 : Using AdaBoost learning algorithm,let our machine learn Strong classifier and combine them linearly
• 5/28 : Debug and tune the algorithms. Experiments and finish write-up.

References

[1] American Kennel Club - Complete List of Breeds, http://www.akc.org/breeds/complete_breed_list.cfm

[2] Aditya Khosla, Nityananda Jayadevaprakash, Bangpeng Yao and Li Fei-Fei. Novel dataset for Fine-Grained Image Categorization. First Workshop on Fine-Grained Visual Categorization (FGVC), IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2011.

[3] Stanford Dogs Dataset, http://vision.stanford.edu/aditya86/ImageNetDogs/