Salient Object Detection Based On Local Feature Preservation

Image saliency detection is one of the key technologies in the field of computer vision.It also plays an important role in actual computer vision tasks.Its main goal is to simulate the human visual attention system to find areas of interest automatically.Therefore,the core problems what saliency detection research have to face is: firstly how to obtain detection results consistent with people's subjective will,and secondly how to discriminate the difference between the foregrounds object and background area when the targets of interest are irregular and ambiguous in complex semantic scenarios.In recent years,although deep learning has made progressive developments and achievements in computer vision,the detailed features still are diluted during the top-down propagation process,and there still exists many deficiencies in the aspect of preserving local features.Therefore,in this paper we will proceed from the effective attention,the local detail information to conduct in-depth research on the key issues in saliency detection.The main research work and contributions are as follows:(1)We proposed a multi-scale convolution combined deconvolution network model.More specifically,we applied the deconvolution to the feature layers as well as their contract features,so that more multi-scale parameters could be preserved;then the fusion of the deconvolution offsets were combined with global information to get the salient result.The experimental results show that with many uncertainty factors in the complex background,compared to the traditional methods,the proposed method could get a satisfactory salient detection,and compared to the up-to-date deep learning methods,our method can produce more clear and accurate areas,which reduces the loss of information to some extent and restores more details,at the same time,the runtime of our method has been accelerated due to the design of the independence between thedeconvolution layers.(2)We proposed a saliency detection algorithm based on multi-group dilated convolution network.In order to effectively expand the receptive field to obtain richer and more effective contextual semantic information,and to reduce the loss of most useful features,the multi-scale context-aware feature extraction module was constructed by using the grouping expansion convolution layer with multiple expansion rates,which could learn the target area information and the semantic information in image context;this feature extraction module mainly focused on the effective high-level context features and low-level spatial structure features.Then the contrast feature was used to capture the difference of each feature relative to the light or dark area around it,and the combination of the convolution and deconvolution model was used to obtain the multi-layer local detail information.Finally,the predicted results were effectively fused to obtain the final significance graph.The quantitative and qualitative experiments on five benchmark data sets show that the method in this part performs better than many saliency detection methods,and is still more successful in complex scenarios with many uncertain factors,such as irregular targets and ambiguity.Our method has good self-adaptability,can be applied to various scene changes,and improves the performance of the saliency detection algorithm to a certain extent.In summary,we combines a multi-scale deconvolution network model and a multi-group dilated convolutional network model and applies it to image saliency detection tasks.The proposed methods,in a large amount of experimental results,observably illustrate the effectiveness of the proposed model,and also provide some new views for the future study of image saliency detection algorithms.