Saliency detection has been increasingly gaining research interest in recent years since many computer vision applications need to derive object attentions from images in the first steps. Multi-scale awareness of the saliency detector becomes essential to find thin and small attention regions as well as keeping high-level semantics. In this paper, we propose a novel holistic and deep feature pyramid neural network architecture that can leverage multi-scale semantics in feature encoding stage and saliency region prediction (decoding) stage. In the encoding stage, we exploit multi-scale and pyramidal hierarchy of feature maps via the densely connected network with variable-size dilated convolutions as well as a pyramid pooling. In the decoding stage, we fuse multi-level feature maps via up-sampling and convolution. In addition, we utilize the multi-level deep supervision via plugging in loss functions at every feature fusion level. Multi-loss supervision regularizes weights searching space among different tasks minimizing overfitting and enhances gradient signal during backpropagation, and thus enables us training the network from scratch. This architecture builds an inherent multi-level semantic pyramidal feature maps at different scales and enhances model’s capability in the saliency detection task. We validated our approach on six benchmark datasets and compared with