Genetic algorithms based feature combination for salient object detection, for autonomously identified image domain types

Combining features from different modalities and domains has been demonstrated to enhance the performance of saliency prediction algorithms. Different feature combinations are often suited to different types of images, but existing techniques attempt to apply a single feature combination across all image types. Furthermore, existing normalization and integration schemes are not utilized in salient object detection as the combination of potential solutions is intractable to test. The aim of this work is to autonomously learn feature combinations for autonomously identified image types. To this end, we learn optimal normalization and integration schemes along with feature weightings using a novel Genetic Algorithm (GA) method. Moreover, we learn multiple image dependent parameters using our novel image-based GA (IGA) approach, to increase the generalization of the system on unseen test images. We present a thorough quantitative and qualitative comparison of our proposed methods with the state-of-the-art benchmark and deterministic methods on two difficult datasets (SED1 and SED2) from the segmentation evaluation database. IGA shows superior performance through learning optimal parameters depending upon the composition of images and using feature combinations appropriately enhances test performance and generalization of the system.