Scene and Object Recognition

Title scene recognition using global properties
Abstract One of the biggest mysteries of the vision is human's brain remarkable ability to understand quickly and effortlessly new scenes, places and events. Compared to object recognition, scene recognition is more challenging, due to the ambiguity and variability in the content of the scene images, mostly due to lighting changes and scale effects that seems bad. A scene is the place that we can move in it. To distinguish scene from ’object’ or ’texture’, we consider the absolute distance between observer and the fixated zone as the discriminating factor. So, an ’object’ is something that subtends about 1 to 2 meters around the observer; but in case of a scene, the distance between the observer and the fixated point is usually more than 5 meters. Scene recognition methods use from local features that don’t consider spatial relations between scene regions and global features of the images. Methods based on global features consider whole of scene as a one thing. Here, we extract gist and spectrum signatures features from 8 outdoor scene categories and combine them to recognize images of dataset by a support vector machine classifier with a gaussian kernel.
Supervisor Prof. Jamshid Shanbehzadeh
Advisor Dr. Zeinab Ghassabi
Student Fatemeh Ghanbari Adivi, MSC student of Computer Engineering, Artificial Intelligence, Khawrazmi University of Teheran
Title Contexual Based Indoor Scene Classification
Abstract Understanding the meanings and the contents of images remains one of the most challenging problems in machine intelligence an statistical learning. Although low-level image features have been proven to be effective representations for a variety of visual recognition task, such as object recognition and scene classification, but in fact that low-level image features carry little semantic meanings. Instead using high-level image features carry more semantic meanings and can represent better result in some field specially scene classification. In this thesis we study indoor scene classification based on contexual features.
Supervisor Prof. Jamshid Shanbehzadeh
Advisor Dr. Zeinab Ghassabi
Student Maryam Fany, MSC student of Computer Engineering, Artificial Intelligence, Islamic Azad University, Science and Research Branch, Tehran
Title Indoor Scene Classification Using local features
Abstract Scene classification is highly considered as a complex part of image recovery in machine learning. Feature selection is the most important challenge in scene classification algorithms, because choosing areas which include prominent concepts will increase classification accuracy. {There are two types of image features generally, general features and local features selection./Image features divided in to two types generally, general features and local features selection.}. Sparse and locality-considered linear coding methods are the most important feature extraction methods which are based on choosing image areas. Choosing areas which include important information will cause to extract the most important and the best input features that using them will improve the scene classification.
Supervisor Prof. Jamshid Shanbehzadeh
Advisor Dr. Zeinab Ghassabi
Student Mohabbat Hajati, MSC student of Computer Engineering, Artificial Intelligence, Islamic Azad University, Science and Research Branch, Tehran
Title Context-based object recognition in digital images
Abstract Object recognition systems are an important and crucial part of modern intelligent systems. In real world there is strong relation between objects and environment. When a special object is searched in a scene, spectator will focus on locations with highest prior probability for object existence. Therefore scene context is very important for decision making about eye movement. Context can be classified into three category: semantic context, spatial context and scale context. Object recognition models use contextual information from a local and global image level. Global context considers image statistics from image as a whole and local context considers context information from neighboring areas of the object.
Supervisor Prof. Jamshid Shanbehzadeh
Advisor Dr. Zeinab Ghassabi
Student Soheila Sheikh Bahaei, MSC student of Computer Engineering, Artificial Intelligence, Khawrazmi University of Teheran