Downloads & Free Reading Options - Results

Image compression usually considers the minimization of storage space as its main objective. It is desirable, however, to code images so that we have the ability to process the resulting representation directly. In this thesis we explore an approach to document image compression that is efficient in both space (storage requirement) and time (processing flexibility). A representation is presented in which component-level redundancy is removed by forming a prototype library and component location table. This representation forms a basis for compression and provides direct access to image components. To generate the prototype library, a new clustering approach is developed which is suitable for document image components. The distance metric is based on a character degradation model so that degraded versions of the same character will be grouped together. To achieve a lossless representation when required, the residuals are encoded efficiently using a structural distance ordering. OCR is then used as a measure of readability to evaluate the rate distortion tradeoff for lossy compression. A set of algorithms is presented for typical document processing applications which operate effectively on the compressed representation.

Contents:What is a Document Image and What Do We Do With It?; Preparing the Document Image; Finding Appropriate Features; Recognizing Components of Text Documents; Recognizing Components of Graphics Documents. Lecture Notes Collection FreeScience.info ID2079 Obtained from http://www.cse.usf.edu/~r1k/DocumentImageAnalysis/DIA.pdf http://www.freescience.info/go.php?pagename=books&id=2079

Includes bibliographical references (p. 491-523)

In this paper, we present a novel approach for initializing deep neural networks, i.e., by turning PCA into neural layers. Usually, the initialization of the weights of a deep neural network is done in one of the three following ways: 1) with random values, 2) layer-wise, usually as Deep Belief Network or as auto-encoder, and 3) re-use of layers from another network (transfer learning). Therefore, typically, many training epochs are needed before meaningful weights are learned, or a rather similar dataset is required for seeding a fine-tuning of transfer learning. In this paper, we describe how to turn a PCA into an auto-encoder, by generating an encoder layer of the PCA parameters and furthermore adding a decoding layer. We analyze the initialization technique on real documents. First, we show that a PCA-based initialization is quick and leads to a very stable initialization. Furthermore, for the task of layout analysis we investigate the effectiveness of PCA-based initialization and show that it outperforms state-of-the-art random weight initialization methods.

In this paper, we present a novel approach for initializing deep neural networks, i.e., by turning PCA into neural layers. Usually, the initialization of the weights of a deep neural network is done in one of the three following ways: 1) with random values, 2) layer-wise, usually as Deep Belief Network or as auto-encoder, and 3) re-use of layers from another network (transfer learning). Therefore, typically, many training epochs are needed before meaningful weights are learned, or a rather similar dataset is required for seeding a fine-tuning of transfer learning. In this paper, we describe how to turn a PCA into an auto-encoder, by generating an encoder layer of the PCA parameters and furthermore adding a decoding layer. We analyze the initialization technique on real documents. First, we show that a PCA-based initialization is quick and leads to a very stable initialization. Furthermore, for the task of layout analysis we investigate the effectiveness of PCA-based initialization and show that it outperforms state-of-the-art random weight initialization methods.

In this paper, we present a novel approach for initializing deep neural networks, i.e., by turning PCA into neural layers. Usually, the initialization of the weights of a deep neural network is done in one of the three following ways: 1) with random values, 2) layer-wise, usually as Deep Belief Network or as auto-encoder, and 3) re-use of layers from another network (transfer learning). Therefore, typically, many training epochs are needed before meaningful weights are learned, or a rather similar dataset is required for seeding a fine-tuning of transfer learning. In this paper, we describe how to turn a PCA into an auto-encoder, by generating an encoder layer of the PCA parameters and furthermore adding a decoding layer. We analyze the initialization technique on real documents. First, we show that a PCA-based initialization is quick and leads to a very stable initialization. Furthermore, for the task of layout analysis we investigate the effectiveness of PCA-based initialization and show that it outperforms state-of-the-art random weight initialization methods.