Face-recognition-based dog-breed classification using size and position of each local part, and PCA

In face-recognition-based dog-breed classification, dog faces in the same breed are overall different, and dog faces among different breeds are similar. This leads to misclassification. To classify dog breed using face local parts which represent the breed characters can improve the accuracy. However, finding and capturing each local part from dog face images are time-consuming. This paper proposes a method to improve speed of dog breed classification using size and position from local parts in the dog face images and PCA. Firstly, size and position of each local part form the dog face images is trained. Then, the sizes and positions are used to find and capture each local part from test images. Within each part, the Principle Component Analysis (PCA) is applied to classify the dog breed. In the PCA-based classification, Test local part images are represented in term of Eigen vectors. The vectors are compared with main feature part templates for each breed in the database and are calculated as weight. The image under test is classified as the breed that gives the minimum weight between test image and train image. To evaluate the performance of the proposed method, experiments with 350 dog face images from 35 dog breeds had been performed. Before the testing, 3 dog face images from every breed (totally 105 dog faces) are divided into 3 main feature parts (left ear, right ear, face without ears) to train the system. The experiments show that he proposed system is faster than the conventional system around 4 times.