[Week-3 Clean/Messy Rooms]
Members : Atakan Erdoğdu, Damla Ünal, Zekeriya Onur Yakışkan, Didem Yanıktepe
Previous weeks, we have introduced our dataset to you. This week, we renamed the images in the data in each class with respect to the format “<number>.jpg”. The numbers start from 0 to the class size.
We reviewed all the photos we’ve got. We deleted inappropriate photos or moved them to appropriate classes.
This week we added messy room photos to our data. We took these photos from Google. However, the number of photos for messy rooms was not enough. We are currently trying to increase our data set for messy rooms with data augmentation techniques.
As a result of our conversation with Necva Bölücü, we gave up the idea of classifying 6 classes in one step. Instead, we combine messy and clean room photos and find the room name for the 3 classes(living room, kitchen, bedroom) first. After that, we will try to classify rooms as messy or clean. We may use the room predictions while classifying rooms as messy or clean. In this way, we aim to reduce the margin of error and to use our code in 2 different ways.
Data Augmentation
In machine learning, more data would lead to a better model. Therefore, the lack of available data in applications may decrease the success of the model. In order to prevent this situation, synthetic data is produced from the current data set in order to increase the number of samples. This is called “Data Augmentation”.
The purpose of producing synthetic data is making data sets more diverse as in the real world. In object recognition problems, there are some techniques for data augmentation such as picture rotation, translation, scaling, adding noise to the image (Gauss, etc.), cropping the picture from a specific region… In this project, we will enlarge our dataset by changing the lighting conditions of pictures.
The lighting situation: This is a very important variety in the dataset, not only for the network to properly learn about the object, but also for simulating the practical scenario of images taken by the user. The lighting condition of the images changes by adding Gaussian noise to the image.
This approach leads to a very important variety in the dataset. It is for simulating the practical scenario of images taken by the user. The lighting condition of the images changes by adding Gaussian noise to the image.
Types of classification algorithms we thought about our project
1.Logistic Regression
Logistic regression is a broadly used statistical classification model.Logistic Regression could help use predict whether the rooms. Logistic regression predictions are discrete (only specific values or categories are allowed). We can also view probability scores underlying the model’s classifications.
2. Naive Bayes Classifier (Generative Learning Model) :
Classifying image using multinomial naïve bayes is now a common technique, notably. Our goal with this project was to attempt to adopt the same technique to classifying a class of images(e.g. bedroom, kitchen, living room). As presented, a Multinomial Bayes classifier will make effort to use the frequency of room number in pictures according to all classes that intend to represent, to find out labeled images class. A basic assumption and the name of the technique is to think that images are independent of each other, so it sets aside us to create our probability model and use Bayes theorem in parallel with all images in the class.To compute the possibility of an image belonging to a particular class.
3) Convolutional Neural Networks
Sounds like a weird combination of biology and math with a little CS sprinkled in, but these networks have been some of the most influential innovations in the field of computer vision.
Image classification, take an input image or a possibility of classes that best describe the image. For us, the task is one of the first skills we learned from the moment we were born, and as adults it is a normal and easy skill.Without thinking , we can quickly recognize the environment that surrounds us . When we see an image or look at the world around us, often we can characterize the scene immediately and give each object a label without freely noticing it. The ability to recognize patterns quickly, generalize from previous knowledge, and adapt to distinct image environments are skills that we do not share with another machines.
Convolutional Neural Networks is very relevant for image classification, due to fact that CNN not only controls the distance of the reliances to be modeled, yet also enables close input elements to interact with lower layers while remote elements interact with higher layers.CNN can generate ordered abstract representations of images by stacking different convolution layers.
References:
