Research On Image Salient Object Detection Method With Multi-scale Aggregation And Edge Learning

Salient object detection algorithm obtains the most important information in the image,which makes subsequent data analysis more convenient.Salient object detection task extracts the most attractive object area in image and segment the edge of the salient object.Existing salient object detection methods predict better results for pictures of simple scenes.However,it is still difficult to detect pictures with attributes such as low contrast or complex shapes.It is mainly reflected in two aspects.On the one hand,the overall prediction of the internal area of the salient object is not accurate enough.On the other hand,the pixels around the edge of the salient object are more difficult to predict and the edge prediction is not precise enough.Therefore,designing a salient object detection model with precise positioning and fine edges is a key issue.Based on the summary and analysis of the research status in the field of saliency detection,the method in this thesis starts from the perspective of multi-scale feature fusion and edge learning.The main proposed work is as follows:(1)Research on multi-scale aggregation salient object detection model with dilated convolution.The thesis defines a dilated convolution block operator under the VGG16 architecture,which combines dilated convolution and standard convolution to enhance the features extracted by the encoder block.Then,the thesis introduces a deep feedback module to feedback high-level features to low-level features.This method can fully capture global and local information.The thesis uses cross-entropy loss for model training.Multiple comparative simulation experiments are performed on the benchmark datasets.They verified that the proposed multi-scale aggregation method can capture more accurate salient object positions.(2)Research on salient object detection model based on edge learning.The thesis analyzes the edge extraction method with convolutional network and the association with salient object detection.The edge flow is introduced through bidirectional feature pyramid sharing,which improves the multi-scale aggregation model.And the hierarchical edge loss is defined.The output on each side and the final fusion map of the multi-scale aggregation model are provided additional edge loss.A number of comparative simulation experiments are performed on the benchmark datasets.The hierarchical edge loss has a better guiding effect,which makes the model predict more refined salient object edges.