Gated Deep Layer Aggregation For Salient Object Detection

It is the one of main characters for the human visual system to select and simplify the complex information in the scene.Computer vision imitates this character of the human visual system to introduce saliency detection.Salient object detection,as a preprocessing step,aims to obtain the most attracting regions in the image.With the deep learning develop rapidly recent years,a large number of computer vision tasks turned the way to tackle the problem from designing algorithm based on hand-craft to designing deep learning model.It isn't an exception for saliency detection.A lot of methods based on Convolutional Neural Network(CNN)were proposed instead of earlier methods based on heuristic methods.Some of these methods bring a big improvement in performance.However,these methods still exist some problems.For example,some early methods utilized fully connected layer to predict saliency results,which usually results in the loss of spatial information and cost lots of computation resources.The emergence of Fully Convolutional Neural Network(FCN)promoted the advance of end-to-end or pixel-to-pixel tasks including salient object detection.The problem following this is how to utilize multi-level features to obtain high-quality saliency maps.Some methods just used features from the deeper layer or simple way to fuse multi-scale features.These all influence the final prediction results.This paper proposes a gated deep layer aggregation model which is used as saliency detection.The aim is to fully and usefully utilize the context-aware features in the network to acquire high-quality prediction results.Firstly,exploiting VGG16 as multi-level features extraction module obtains multi-scale features which are used in the aggregation stage.Then we introduce a Deep Layer Aggregation(DLA)which is originally used to solve semantic segment.Due to the fact that the human visual system extracts useful information from the scene,which is a crude to fine process.The proposed method takes the aggregation way from the deep layer(location information)to shallow layer(fine information)instead of the original aggregation way.Although this way is able to make the multi-scale features to be fully merged,not any information is helpful for the problem.In this situation,gate function is add to control the information passing in the network during the aggregation period.Gate function allows useful information and prevents useless information to be passed.In theory,this paper designs a module which is able to fully and usefully merge multi-level features.Finally,the side-outputs from different aggregation modules are merged as final outputs.During the training process,two public datasets are utilized as training datasets.To get more training data,using mirror images and rotated images augments datasets to increase the capable of generalization ability.For verifying the performance of proposed method,this paper fully evaluates the proposed method on four public datasets.The method of this paper compares with 14 state-of-the-art works.These methods include different types of saliency detection method including heuristic methods,traditional machine learning methods and CNN-based methods.The proposed method shows outstanding performance under different evaluating indicators.