Andrea Kutics

Self-adaptive Feature Extraction Scheme for Mobile Image Retrieval of Flowers

This paper proposes a new self-adaptive feature extraction scheme to improve retrieval precision for Content-based Image Retrieval (CBIR) systems on mobile phones such that users can search similar pictures for a query image taken from their mobile phones. The proposed methods employ a newly modified extraction method using the Canny edge-based Edge Histogram Descriptor (CEHD), Color Layout Descriptor (CLD) and the Curvature Scale Space (CSS) shape-based descriptor. To obtain object shapes, salient regions are detected by means of a multi-scale self-developed segmentation model. Experiments were conducted using flower images as image data in order to verify the most pertinent feature extraction methods in designing a domain knowledge-driven self-adaptive feature extraction scheme. Test results prove that the CSS descriptor is useful to determine prominent features of a flower image before employing additional extraction techniques. By that means, the system can enhance retrieval precision and avoid unnecessarily extracting insignificant features.

MOSIR: image and segment-based retrieval for mobile phones

This paper proposes a novel image retrieval system called MOSIR, developed for mobile phones. This system enables the user to find similar images and information related to the photos taken with the users mobile phone at any time. It overcomes the problem originating from small display panels and limited control keys. Similar images are then retrieved by extracting edge-based and color-layout features. Region-based queries are also processed by detecting salient regions and extracting their features. The system uses both self-defined email and web application software to provide flexible image searching, and it overcomes security problems. A new user interface is designed in Flash Lite to display the results in different sizes and contexts. The method provides satisfactory results. This mobile system proved to be useful for finding images and semantically connected information while the user was on the move.

Layered Self-Organizing Map for Image Classification in Unrestricted Domains

The inherent difficultly in unrestricted image domain classification is due to the many different features exhibited by images. Efforts made toward classification of abstract features tend to focus on a single attribute. Without a method of unifying descriptors, it becomes very difficult to perform multi-feature analysis. Extending the concept of the Self-Organizing Feature Map to include multiple competitive layers, it has been possible to create a new type of Artificial Neural Network capable of analyzing image and signal datasets with multiple feature descriptors concurrently in a powerful yet computationally light manner. Compared to standard CBIR retrieval approach, a marked increase in the precision of clustering of 13 points has been achieved, along with a reduction in computation time.

Parallel Method for Automatic Shape Determination via the Evolution of Morphological Sequences

A parallel evolutionary method for object shape determination is proposed by automatically generating morphological operator and operation sequences. Artificial individuals built up from binary morphological operators and operations undergo recombination and mutation processes for producing new generations. The normalized correlation between the generated shape and the corresponding input image region is calculated for fitness. This method requires no preliminary knowledge of the object shape and also no constraints are used for image background and smoothness. The parallel evolutionary approach provides a fast and directed search on large number of possible morphological sequences and the method can be applied on a wide range of images. The morphological operations are implemented by low level image processing steps and executed as parallel tasks by applying both data and algorithmic parallelization. As a concrete application, this method is utilized for the shape determination of skin objects in a system consisting of a camera device connected to a grid architecture of transputer nodes.

Guest editorial: online multimedia signal and image processing

In recent years, multimedia has become an essential part of networking and online communication environments. As people connect to the Internet or to one another multimedia content such as photos, video and music, takes the lion’s share of the overall artwork traffic. Multimedia-based applications use a wide range of signal and image processing techniques to enhance user experience. Semantic multi-media retrieval has been a major research focus for the last decade. Recently, a large number of cutting-edge algorithms have been developed performing audio, image and video content extraction and object recognition. Artificial intelligence and machine learning techniques have been also extensively used to connect contextual and conceptual information to multimedia data. The importance of metadata has been boosted by the emergence of high-end devices such as tablets, smartphones, and other hand-held devices. Sematic-aware metadata can further improve intelligence of online multimedia environments. This special issue targets online applications ranging from traditional signal and imageprocessing methods to evolving semantic-based multimedia, audio, image and video technologies. Also, it aims to demonstrate novel, flexible and easy-to-use applications developed for the next generation of devices. This special issue received many submissions from researchers and practitioners working on multimedia signal and image processing. After two rounds of review, eventually seven high quality papers were chosen. The paper titled “Extending the Image Ray Transform for Shape Detection and Extraction” presents a novel approach to image analysis performing feature extraction at low level and complements it with high-level feature extraction to determine structure, exploiting the Hough transform. Authors analyze performance with images from the Caltech-256 dataset and describe how thir approach can select chosen shapes.

Simple Example: Clustering Images Using Expectation Maximization

This paper gives an example of a novel simple implementation of the EM algorithm for clustering images. Here we use a simple gray scale color feature to describe an image. When compared to results of other methods using the same simple feature, we found that the proposed method performs well. These comparison results imply that this simple model can be extended to cluster images using more complex features such as texture, shape, and other color descriptors to further improve the precision and recall of the results in order to outperform the existing methods. Further research can prove that this simplified EM algorithm achieves robust classification of unrestricted image domain.

Segment-based image classifcaton using Layered-SOM

Arbitrary domains represent one of the most difficult areas for image classification algorithms to categorize effectively. Inconsistent features require a computationally expensive multipartite approach to search for possible underlying structures within datasets. This paper proposes a new approach to the problem by applying a self-developed, non-linear, multi-scale image segmentation method to identify and extract prominent regions among several visual features expressing color, texture and layout properties. Integrating this method with the Layered Self-Organizing Map has achieved a simple yet powerful multifaceted Artificial Neural Network classifier for mixed domains which has improved abstract classification precision when compared against unsegmented classification methods.

Mapping image segments to words for image retrieval

This paper proposes a new method for mapping image segments to words in three layers for image retrieval. Our main goal here is to incorporate higher-level semantics into the retrieval process and thus narrow the gap between the users interpretation and the automatically extracted low-level visual features of the same image content. The method is based on nonlinear segmentation, as well as clustering and statistical learning applied to both visual and textual features to find semantic relations between visual segment clusters and words of various abstraction levels. Experiments conducted on a wide, natural image domain shows that step-by-step semantic inferencing in image-word mapping helps to improve retrieval performance. The method supports various textual and/or visual browsing and searching schemes and is proved to be very useful for effective browsing and retrieval in large image data sets.

A New Parallel Method Based on a Genetic Approach for Determination and Classification of Skin Spots

A low cost CCD camera system connected to a transputer network as a parallel processing device has been developed for the determination of human skin objects. Ultraviolet, visible and penetrative infrared images recorded by CCD cameras are used as the input data. A new genetic method based on mathematical morphology has been developed to detect objects and estimate shape properties of skin objects such as speckles and blood vessels on an arbitrarily chosen area. A very fast system could be achieved, and as no presumptions are used on the shape of the objects in question, the developed method is widely applicable to various kinds of images. The described transputer network was found to be very suitable for object analysis tasks due to its high performance and flexibility.