Overview of Dog Breed Classification using various technologies

Affiliation: Shruti Shah, Fourth Year B. Tech Integrated, Computer engineering, NMIMS's MPSTME.

Pattern recognition (PR) is understood as a human recognition process that can be completed by computer technology. We should add information for identifying some object into the computer. For doing the same, we need to abstract the object which must be recognized and create its mathematical model to describe it. Then we replace the object concerning what the machine can take and process. In simple words, the pattern recognition system is to identify the category to which the object belongs, such as the face in face recognition. This article is based on a pattern recognition system which will help us to identify the dog breed using various technologies.

This classification uses computer vision, machine learning, and deep learning techniques to predict the breed by just the images. The dog facial key points are recognized in each image using a convolutional neural network. These key points are used to extract features via SIFT descriptors and histograms. After this, we compare a variety of classification algorithms, which will help us in predicting the breed of dogs.

This project is aiming to classify photos of dog breeds. There are very few variations in scale, shape, and color between the breeds and large variation within the breeds. Owing to the complexity, the challenges of recognizing the breeds are impaired by the variations in photos used at stud, showing dogs of the same breed in several lights and positions. 

Objective

The main objective is to use deep learning algorithms and to predict the breed of different dogs. We will look at a trained model that is used on 133 breeds of dogs on over 1 lakh images also. CNN will play an important role in recognizing the patterns in the training data and to make changes to the model. This model will be trained to guess the breed of the dog even if it is a puppy.

Datasets

We will use the Columbia Dogs Dataset for our tests, as it provided the most reliable online data available. The dataset contains over 8000 images of 130+ different dog breeds some of which are featured. Each image had a corresponding text file that annotated both bounding boxes and key points for each dog face. The facial key points annotated were the right eye, left eye, nose, right ear tip, right ear base, head top, left ear base, and left ear tip, which we used to train for key-point detection in our convolutional neural network.

Key point detection

The first phase of the study is key-point identification to better classify dog breeds. In the Columbia dataset, dog face key points are identified and annotated as the right eye, left eye, nose, right ear tip, right ear base, head top, left ear base, and left ear tip. Given the unseen image from the test sample, estimate the dog's face key points in terms of pixels as close to the ground truth points as possible.

Feature Extractions

When the facial key points were found, we used them to extract relevant information about the picture that we would use later in the classification. SIFT Descriptors One function used was centered grayscale SIFT descriptors at important key points. SIFT descriptors do an amazing job describing located regions of an image in such a way that it can be contrasted later with located parts of other images (e.g. we can compare one dog's eye to another to see if they look similar). SIFT descriptors can be based on the left and right eyes, the nose, and the facial center (calculated as the center of these three points). Sift descriptors can be rotated to suit the dogs face rotation (calculated for the line connecting the two eyes as the rotation away from horizontal). The descriptors were also designed to be half the distance between the two heads.

Conclusion

In the end, we concluded that the deep learning model has a very great capability to surpass the human potential if the data provided is sufficient. In the future, deep learning will create another deep learning model on its own and deep learning models will write codes and surpass human capabilities. Deep learning has a lot of scope in medical sciences by analyzing the images from deep convolution neural network. Deep learning may be one of the possible reasons for the destruction of humankind. Dog breed classifier is one of the mini-projects of deep learning developed using the exception model and advanced neural networks. Transfer learning has a great scope in the future by combining a prebuilt model with the model we constructed.

Sources:

?https://medium.com/@rojandhimal1/cnn-and-transfer-learning-for-dog-breed-classification-34de15596cdc

? M.D.Zeiler, R. Fergus, "Visualizing and understanding convolution neural network ", ECCV, 2014.

? D. Nouri. Using convolutional neural nets to detect facial key points tutorial, 2014.

? https://towardsdatascience.com/dog-breed-classification-hands-on-approach-b5e4f88c333e

? www.stanford.edu