Adrian Carballal

Computerized measures of visual complexity.

Visual complexity influences peoples perception of, preference for, and behaviour toward many classes of objects, from artworks to web pages. The ability to predict peoples impression of the complexity of different kinds of visual stimuli holds, therefore, great potential for many domains, basic and applied. Here we use edge detection operations and several image metrics based on image compression error and Zipfs law to estimate the visual complexity of images. The experiments involved 800 images, each previously rated by thirty participants on perceived complexity. In a first set of experiments we analysed the correlation of individual features with the average human response, obtaining correlations up to rs = .771. In a second set of experiments we employed Machine Learning techniques to predict the average visual complexity score attributed by humans to each stimuli. The best configurations obtained a correlation of rs = .832. The average prediction error of the Machine Learning system over the set of all stimuli was .096 in a normalized 0 to 1 interval, showing that it is possible to predict, with high accuracy human responses. Overall, edge density and compression error were the strongest predictors of human complexity ratings.

Using complexity estimates in aesthetic image classification

In recent years, the search for computational systems that classify images based on aesthetic properties has gained momentum. Such systems have a wide range of potential applications, including image search, organization, acquisition and generation. This work explores the use of complexity estimates to predict the aesthetic merit of photographs. We use a set of image metrics and two different classifiers. Our approach classifies images gathered from a photography web site, attempting to reproduce the evaluation made by a group of users. For this purpose, we use complexity estimate metrics based on the encoding size and compression error of JPEG and fractal compression, which are applied to the original value channel and to the images resulting from applying Sobel and Canny filters to this channel. By employing these estimates, in conjunction with the average and standard deviation of the value channel, i.e., 20 features, a success rate of 74.59% was attained. Using the three most influential features yields a success rate of 71.34%, which is competitive with the best results reported in the literature, 71.44%, using the same dataset.

Aesthetic classification and sorting based on image compression

One of the problems in evolutionary art is the lack of robust fitness functions. This work explores the use of image compression estimates to predict the aesthetic merit of images. The metrics proposed estimate the complexity of an image by means of JPEG and Fractal compression. The success rate achieved is 72.43% in aesthetic classification tasks of a problem belonging to the state of the art. Finally, the behavior of the system is shown in an image sorting task based on aesthetic criteria.

Computing Aesthetics with Image Judgement Systems

The ability of human or artificial agents to evaluate their works, as well as the works of others, is an important aspect of creative behaviour, possibly even a requirement. In artistic fields such as visual arts and music, this evaluation capacity relies, at least partially, on aesthetic judgement. This chapter analyses issues regarding the development of computational systems that perform aesthetic judgements focusing on their validation. We present several alternatives, as follows: the use of psychological tests related to aesthetic judgement; the testing of these systems in style recognition tasks; and the assessment of the system’s ability to predict the users’ valuations or the popularity of a given work. An adaptive system is presented and its performance assessed using the above-mentioned validation methodologies.

Feature selection and novelty in computational aesthetics

An approach for exploring novelty in expression-based evolutionary art systems is presented. The framework is composed of a feature extractor, a classifier, an evolutionary engine and a supervisor. The evolutionary engine exploits shortcomings of the classifier, generating misclassified instances. These instances update the training set and the classifier is re-trained. This iterative process forces the evolutionary algorithm to explore new paths leading to the creation of novel imagery. The experiments presented and analyzed herein explore different feature selection methods and indicate the validity of the approach.

Guest editorial: special issue on biologically inspired music, sound, art and design

The idea of using evolutionary computation for artistic purposes can be traced back to Dawkins. In his 1987 book, The Blind Watchmaker [1], Dawkins presents a program that allows the evolution of the morphology of ‘‘virtual creatures’’. The user guides the genetic algorithm, indicating the favorite biomorphs, avoiding the need to develop a fitness function. Using a similar approach, the seminal works of Sims [2], where genetic programming was used to evolve populations of images, and of Todd and Latham [3], where evolutionary techniques are used to evolve 3D forms with organic appearance, led to the emergence of a new research area, evolutionary art, which is characterized by the use of nature-inspired computing for artistic purposes. At the same time, early works in the field of evolutionary music explored the development of hard-wired functions [4] and of artificial neural networks [5] to assign fitness. Soon afterwards, Baluja et al. [6] applied artificial neural networks to guide the evolution of images. As is often the case in new research areas, particularly interdisciplinary ones, the first years were characterized by individual efforts and uncoordinated research, which often led to the reinvention of the wheel. This problem was amplified by the lack of scientific events specifically dedicated to the area, which led to the publication of these early efforts in generic evolutionary computation, artificial intelligence, neural networks or computer music conferences and journals.

Machine Learning in Biomedical Informatics

Biomedical informatics has skyrocketed in the last years by reducing sequencing costs with next-generation sequencing techniques. Thus, the amount of available data to study is increasing excessively, and more recently, is even open access for researchers. Due to this, biomedical informatics researchers, with different profiles, are using machine learning algorithms for knowledge extraction and, despite the great amount of benefits this entails, it also requires to take into account a series of particularities of inexcusable compliance in order to achieve a solution which is real.

Authorship and Aesthetics Experiments: Comparison of Results between Human and Computational Systems

This paper presents the results of two experiments comparing the functioning of a computational system and a group of humans when performing tasks related to art and aesthetics. The first experiment consists of the identification of a painting, while the second one uses the Maitland Graves’s aesthetic appreciation test. The proposed system employs a series of metrics based on complexity estimators and low level features. These metrics feed a learning system using neural networks. The computational approach achieves similar results to those achieved by humans, thus suggesting that the system captures some of the artistic style and aesthetics features which are relevant to the experiments performed.

Distinguishing paintings from photographs by complexity estimates

This study is aimed at exploring the ability of complexity-based metrics to distinguish between paintings and photographs. The proposed features resort to edge detection, compression and entropy estimate methods that are highly correlated with artwork complexity. Artificial neural networks based on these features were trained for this task. The relevance of various combinations of these complexity metrics is also analyzed. The results of the current study indicate that different estimates related to image complexity achieve better results than state-of-the-art feature sets based on color, texture and perceptual edges. The classification success rate achieved is 94.82% on a dataset of 5235 images.

Automatic multiscale vascular image segmentation algorithm for coronary angiography

Abstract Cardiovascular diseases, particularly severe stenosis, are the main cause of death in the western world. The primary method of diagnosis, considered to be the standard in the detection and quantification of stenotic lesions, is a coronary angiography. This article proposes a new automatic multiscale segmentation algorithm for the study of coronary trees that offers results comparable to the best existing semi-automatic method. According to the state-of-the-art, a representative number of coronary angiography images that ensures the generalisation capacity of the algorithm has been used. All these images were selected by clinics from an Haemodynamics Unit. An exhaustive statistical analysis was performed in terms of sensitivity, specificity and Jaccard. Algorithm improvements imply that the clinician can perform tests on the patient and, bypassing the images through the system, can verify, in that moment, the intervention of existing differences in a coronary tree from a previous test, in such a way that it could change its clinical intra-intervention criteria.