A genetic programming framework in the automatic design of combination models for salient object detection

Abstract

In computer vision, the salient object detection problem consists of finding the most attention-grabbing objects in images. In the last years, many researchers have proposed salient object detection algorithms to address this problem. However, most of the algorithms perform well only on images with specific conditions and they do not solve the general problem. To cope with a more significant number of image types than those where each standalone saliency detection method performs well, novel methods search to generate a combination model that improves the overall performance of detecting salient objects in images. The contribution of this work is oriented towards the automatic design of combination models by using genetic programming. The proposed approach automatically selects the algorithms to be combined and the combination operators that result in an improvement in the overall performance. The evolutionary approach uses as input a set of candidate saliency detection methods and a set of combination operators. The set of input saliency detection methods includes algorithms from the state-of-the-art. The set of combination operators includes fuzzy logic combination rules, morphological operations, and image processing filters. The outcome of each run of the evolutionary process is a combination model that describes how the input models have to be combined. An advantage of the proposed approach is that these models explain and give insight about which standalone methods are important to improve the response in the solution of the saliency detection problem. The improvement of the final combination models is demonstrated by comparing their performance against that of several state-of-the-art saliency detection methods, that of several classic combination models, and that of other evolutionary computation-based approaches, on four benchmark datasets. The results were analyzed using two statistical tests, the Wilcoxon rank-sum test, and the t-test. Both tests confirmed that the proposed approach outperforms all of the other algorithms under test and that its performance advantage is statistically significant.