Fast salient object detection based on multi-scale feature aggression

Abstract

Since the current CNN-based salient object detection algorithms are of slow speed, as well as fail to preserve rich information of object boundaries, which makes the regions along object contours blurred and inaccurate, a fast salient object detection algorithm with multi-scaled features aggression was proposed. Based on deep ResNet-50, four kinds of features in various resolution levels are extracted separately and then aggregated, which can make the output of network preserve more detailed information about object boundaries, so that the blurry salient maps can be solved via the proposed method. We trained an end-to-end model on THUS10K database, the resulting network can produce a saliency map with pixel-level accuracy from an input image. Extensive experiments on PascalS and DUTOMRON confirm that the proposed method achieves higher AUC and F-measure value while processing images at a rate of 15 fps, which is dramatically faster than any other eight existing methods.