Research On Salient Object Detection Method Based On Attention Recurrent Network

The human visual system can quickly and accurately obtain valuable salient information in certain specific regions of images.In practical applications,researchers attempt to design a saliency detection algorithm that can automatically capture saliency information in images just like the human visual system.At present,salient object detection has become a hot topic in computer vision,which is widely used in image segmentation,object tracking,image retrieval,etc.Benefiting from the powerful feature extraction capabilities and feature expression capabilities of deep convolutional neural networks,the performance of salient object detection methods based on deep learning has been greatly improved over traditional methods.But most of saliency detection methods are still difficult to accurately detect the salient object in complex images.The results may be incomplete,or the boundary of detected salient objects may be blurred.To solve the above problems,this thesis studies the saliency detection based on the attention mechanism and recurrent network.The main contributions are divided into the following two aspects:The feature expression ability of deep neural networks is the key to salient object detection.Convolutional features extracted by most of existing methods lack capability of discrimination.For this reason,we propose a salient object detection method based on multi-feature attention recurrent network.The multi-scale feature extraction module uses the atrous convolution with four different receptive fields to obtain high-level multi-scale context features,and an aggregation operation realizes the fusion of multi-scale features.In order to make the fused feature more discriminative,an attention mechanism is adopted to adaptively increase the weight of each feature position to distinguish their importance.On this basis,the recurrent network is utilized to fine-tune the salient object,which can make salient regions more complete and edges more refined,thereby generating an accurate saliency map.Compared with eight state-of-the-art methods on five public datasets,the saliency maps generated by our method are more accurate in terms of MAE and F-measure.Since the fully convolutional network can extract rich multi-level and multi-scale features,it plays an important role in salient object detection.However,most of existing models based on fully convolutional network use multi-level features in a single way to generate saliency maps that are not accurate enough.To address this issue,we propose a salient object detection method using recurrent guidance network with hierarchical attention features.Firstly,the multi-level features extracted from fully convolutional network are divided into low-level features and high-level features.The multi-scale features are extracted from the high-level features using atrous convolutions with different receptive fields to obtain rich context information.At the same time,the low-level features are refined with a feature refinement module to supplement the detailed information in convolutional features.Due to the different representations of hierarchical features,their attentions are quite different.Therefore,a two-stage attention module is introduced for hierarchical features to guide the generation of saliency maps.After the low-level features and the high-level features are fused,the attention of the model may be biased,leading to deviations in detected salient regions.Hence,a recurrent guidance network is designed to further correct the biased salient regions,which can effectively suppress background interference and progressively refine the boundary of salient objects.The experimental results on the aforementioned five benchmark datasets show that this method has excellent performance in both quantitative and qualitative evaluations.The first method mentioned above pays more attention to the semantic information in high-level features without considering the detailed information in low-level features.In addition,there is no guidance mechanism in the intermediate process of the recurrent network,which generates inaccurate results for some semantically ambiguous images.Furthermore,the second method strengthens the utilization of low-level features to enrich details in convolutional features.Meanwhile,we design a recurrent guidance network using the truth map to supervise the training process of the recurrent network,which can detect the salient objects more accurately.In summary,this thesis demonstrates the importance of attention mechanism and recurrent network in saliency detection,and it provides two feasible saliency detection models for computer vision.