Salient object detection via multi-scale neural network

Abstract

Salient object detection, a fundamental of many computer vision tasks, aims to find the most attractive objects in a given image. In this paper, we propose an end-to-end multi-scale neural network for salient object detection. Firstly, we propose heterogeneous dilated block to effectively increases the receptive field of the network, while alleviating the gridding effect problem caused by dilated convolution. Secondly, we replace the traditional interpolation up-sampling layer with a fully learnable up-sampling module to solve the blurry artifacts and improve the accuracy. Finally, we calculate the loss at three different scales, enabling the network to learn better through back-propagation. The proposed method is validated on MSRA and ECSSD datasets, and shown to outperform the state-of-the-art methods.