Deep Parallel CNN For Multi-scale Saliency Detection Via Global And Local Cue

Saliency detection,which aims to identify the most visually prominent object region in an image,is a very important subject in the field of computer vision and image processing.Serving as a preprocessing step,saliency detection is widely used in computer vision application such as multimedia transmission,image video reconstruction,image video quality evaluation.At the same time,saliency detection also broadly facilitates high-level visual task,such as object detection,identity recognition.As a mature subject,a large number of saliency detection model has been proposed.Traditional saliency detection model can be broadly classified into two categories: hand-crafted model and prior information based method.Hand-crafted model focuses on designing various types of hand-crafted features(e.g.,color,intensity and texture).Though hand-crafted features tend to perform well in standard scenarios,they are not sufficiently robust for complex cases.Prior information based method heuristically enforce some predefined prior on the process of saliency detection such as background prior(regions near image boundaries are probably background).However,it often fails when salient objects touch image boundaries.Motivated by these observations,we need to construct an adaptive model to effectively capture the intrinsic semantic properties of salient objects and their essential differences from the background in a pure-data driven approach.In the same time,the model should detect saliency from both local and global perspective.To solve these issues,deep learning has become a powerful tool in image processing task as pure-data driven approach.Therefore,how to design an effective deep learning model for saliency detection is the focus if this work.This paper design a multi-scale saliency detection model from both local and global view.First,in order to establish an adaptive pure data-driven model,we adapt deep convolutional neural network to extract significant feature from the image automatically.Meanwhile,we utilize superpixel algorithm to simplify computation and strengthen the compact structure and correlation between regions.We also design two deep CNN module to detect saliency from both local and global perspectives.Finally,we add multi-scale information in local deep CNN module to enhance the coherence and consistency of saliency map.The main contribution of this paper are as follows:1)Use superpixel as basic unit for calculation to simplify computation and strengthen the compactness between regions.2)We design a deep CNN model,which contains two modules: CNN-G and CNN-L.CNN-G utilize full image to model saliency from global perspective while CNN-L utilize image patches to model saliency from local perspective.To order to enhance the coherence and consistency of saliency map,a two-way parallel network is used in CNN-L to learn multi-scale feature from two different scale of inputs.3)The experiment show that our model has achieved satisfactory result on three benchmark saliency detection dataset MSRA10 K,DUT-OMRON and ECSSD.Compared to state-of-the-art saliency detection model,our model has achieved better PR-curve and higher F-measure.