Research On Extensibility Of Salient Object Detection Models By Deep Learning

With the development of digital media and the diversification of information dissemination methods,images have gradually become one of the important media for people to receive information.People are exposed to a large amount of image data in life and work,but often only care about a small amount of information in it.Therefore,it is extremely important and meaningful to efficiently and quickly extract information that attracts human attention from images.Salient object detection(SOD)research aims to quickly locate and extract the most attractive objects and regions from images to help people quickly obtain useful information.It is an important starting point for research in computer vision.Influenced by deep learning technology,SOD research has evolved rapidly in the field of computer vision,with significant advances in both effectiveness and efficiency.However,the contradiction between the limited number of training samples in existing SOD datasets and the highly data-dependent feature of deep learning models is still a problem that cannot be ignored,which increases the risk of SOD models overfitting to existing datasets and weakens their generalization.To evaluate the performance of SOD models that perform well on existing datasets in the real visual world,this thesis proposes a model diagnosis method using maximum discrepancy(MAD)competition to examine the generalization of SOD models from another perspective,and further points out the direction of SOD model development and improvement.Furthermore,this thesis proposes a new uncertainty-aware salient object detection model based on the above research.The specific research contents are shown as follows:(1)To evaluate the performance of existing SOD models in the real visual world,this thesis proposes a model diagnosis method using MAD competition to measure the generalization of SOD models.Firstly,a large test dataset(almost 380,000 images)is constructed and a representative sample set(over 3,500 images)is mined from this dataset by a MAD competition method.Then,small-scale subjective experiments are conducted to collect salient object labels and attribute labels for the representative sample set to further validate and analyze the models.Finally,through in-depth analysis of the experimental results,the generalization,advantages and disadvantages of the existing SOD models are clarified,and the direction of SOD model development and improvement is pointed out.(2)Based on the above research,in order to solve the uncertainty problem existing in the SOD models in predicting the salient probabilities of pixels near the contour of salient object,this thesis proposes a new uncertainty-aware salient object detection model.The proposed model adopts internal contour uncertainty map,saliency map and external contour uncertainty map as supervision signals to guide the network to pay attention to the pixels in the salient object and shift its partial attention to the salient and non-salient pixels near the contour of salient object.By doing this,the proposed model can better distinguish "uncertain" pixels near the contour of salient object.In addition,a new feature interaction module is introduced into the model,which aggregates internal contour uncertainty features,saliency features and external contour uncertainty features in the decoding stage to improve the model’s ability to handle "uncertain" pixels.Experimental results show that both of the above operations are helpful to obtain more accurate salient objects and contours.