Salient object detection via hybrid upsampling and hybrid loss computing

Abstract

Salient object detection aims to detect distinct objects which attract human most. It has achieved substantial progress using deep convolutional neural networks in which conventional deconvolution operation is used as recovering the size of image in dense prediction tasks and cross-entropy loss is applied to compute the difference between saliency map and ground truth in pixel level. Different from conventional deconvolution operation, hybrid upsampling block is proposed to retain the detail of object by increasing the receptive field and spatial information when recovering the image size, and hybrid loss which consists of cross-entropy loss and area loss is proposed to train the network optimized by area constraint. At last, an encoder-decoder network based on hybrid upsampling block and hybrid loss is implemented in public benchmark dataset and achieves the best performance against state-of-the-art methods.