Visual Attention Based Salient Object Detection And Its Application

Recently,with the unremitting efforts of scholars,computer vision has developed rapidly,and saliency detection has been widely used in machine learning and image processing,including target recognition,image retrieval,image classification,semantic segmentation and visual question answering.However,with the expansion of image datasets and the increase of scene complexity,traditional salient object detection methods have been difficult to meet the needs of researchers,and it is difficult to capture semantic information in many computer vision tasks(such as edge detection and semantic segmentation),which has great limitations.Recently,due to the rapid development of deep learning technology,salient object detection has made great progress.However,in the follow-up applications of mobile devices,there are still two major challenges:low-resolution output and larger model redundancy.Therefore,this paper presents an accurate yet compact deep saliency detection network.The main research work of this paper are as follows:1.We introduce residual learning into the architecture of HED for salient object detection.Firstly,using the strong semantic capturing ability of the deepest network,the initial saliency map with low resolution is predicted,and then the residual between the original saliency map and the ground truth is predicted sequentially on each side,so as to gradually improve the resolution of the saliency map.The parameters of the network model can be significantly reduced because of learning residual features only.Compared with other existing deep saliency detection networks,this method can not only obtain high resolution saliency map,but also have fewer network parameters and more compact model.2.We further propose reverse attention mechanism to guide side-output residual learning.The main idea is to start with a coarse saliency map generated in the deepest layer with high semantic confidence but low resolution,our proposed approach guides the whole network to sequentially discover missing object regions and details by erasing the current predicted salient regions from side-output features,where the current prediction is upsampled from its deeper layer.Implementing it in a top-down manner,the network can focus on the undetected regions to capture residual details effectively and quickly,which leads to significant performance improvement.3.In order to further reduce the size of the network model,the saliency priori is used as the initial input of the saliency detection network.In this paper,two schemes are designed to generate initial saliency prediction map,namely,learning from Convolution Neural Networks(CNNs)and using existing saliency prior.The experimental results show that,whether these initial predictions are accurate or not,the performance of the proposed method can be improved to the same level as that of the existing best methods by learning the side-output residual based on reverse attention.4.Benefiting from the side-output residual learning unit and the reverse attention mechanism module,this method achieves state-of-the-art performance at present,and with advantages in terms of simplicity,efficiency(45 FPS)and model size(81 MB).By using existing saliency prior map as initial prediction,the model size can be further reduced into 60.7 MB while keep accuracy.The experimental results show that the proposed method is focused on several publicly available datasets has excellent performance in PR(Precision-Recall)curve,F-measure,MAE(Mean Absolute curve Error)and SM(Structure-measure)evaluation index compared with several state-of-the-art methods.In addition,the average execution time is much shorter than the existing methods.Our approach is the best choice for real-time applications until now.In addition,the saliency detection results are also used in background visualization,and good visual effect is achieved.