Text

theses

Salient object detection is a key step in many image analysis tasks such as object detection and image segmentation, as it not only identifies relevant parts of a visual scene but may also reduce computational complexity by filtering out irrelevant segments of the scene. Traditional methods of salient object detection are based on binary classification to determine whether a given pixel or region belongs to a salient object. However, binary classification-based approaches are limited because they ignore the shape of the salient object by assigning a single output value to an input (pixel, patch, or superpixel). In this work, we introduce novel salient object detection methods that consider the shape of the object. We claim that encoding spatial image content to facilitate the information of the object shape can result in more-accurate prediction of the salient object than the traditional binary classification-based approaches. We propose two deep learning-based salient object detection methods to detect the object. The first proposed method combines a shape-preserving saliency prediction driven by a convolutional neural network (CNN) with pre-defined saliency shapes. Our model learns a saliency shape dictionary, which is subsequently used to train a CNN to predict the salient class of a target region and estimate the full, but coarse, saliency map of the target image. The map is then refined using image-specific, low- to mid-level information. In the second method, we explicitly predict the shape of the salient object using a specially designed CNN model. The proposed CNN model facilitates both global and local context of the image to produce better prediction than that obtained by considering only the local information. We train our models with pixel-wise annotated training data. Experimental results show that the proposed methods outperform previous state-of-the-art methods in salient object detection. Next, we propose novel methods to find characteristic landmarks and recognize ancient Roman imperial coins. The Roman coins play an important role in understanding the Roman Empire because they convey rich information about key historical events of the time. Moreover, as large amounts of coins are traded daily over the Internet, it becomes necessary to develop automatic coin recognition systems to prevent illegal trades. Because the coin images do not have the pixel-wise annotations, we use a weakly-supervised approach to discover the characteristic landmarks on the coin images instead of using the previously mentioned models. For this purpose, we first propose a spatial-appearance coin recognition system to visualize the contribution of the image regions on the Roman coins using Fisher vector representation. Next, we formulate an optimization task to discover class-specific salient coin regions using CNNs. Analysis of discovered salient regions confirms that they are largely consistent with human expert annotations. Experimental results show that the proposed methods can effectively recognize the ancient Roman coins as well as successfully identify landmarks in the coin images and in a general fine-grained classification problem. For this research, we have collected new Roman coin datasets in which all coin images are annotated.

ETD_7844

Ph.D.

Includes bibliographical references

by Jongpil Kim

doi:10.7282/T3T15625

ETD doctoral

The author owns the copyright to this work.