Research On Salient Object Detection Based On Multi-modal And Multi-task Learning

Salient object detection refers to the use of mathematical modeling methods to define and segment the most attractive salient area in an image according to the human visual system.At the same time,it can also be regarded as a preprocessing technology to solve the problem of object and background segmentation in other fields.At this time,it is important to study the salient object detection in the clutter.Limited by the amount of single-mode RGB image information,although the performance of salient object detection is still improving,the rate of improvement is decreasing annually and it tends to a saturated state.With the popularization of smartphones and high-performance computing devices,collecting and processing of depth images has become more convenient,which has provided new opportunities for researching salient object detection based on multi-modal.In a cluttered environment,the current mainstream methods cannot show the expected effect.With the help of multi-task learning,combining the two subtasks of salient object detection and salient object existence prediction to solve saliency detection in a cluttered environment is an effective idea.This thesis studies salient object detection methods based on multi-modal and multitask.The main work contents and innovations are summarized as follows:(1)A super supervised RGB-D salient object detection method based on multi-level upsampling fusion is designed.This method uses the upsampling fusion module to provide more cross-modal fusion paths for feature fusion,and combined with the super supervised residual module to reduces the vagueness of multi-modal feature fusion and enhances the adequacy.(2)A salient object detection method based on salient object existence prediction is designed.This method consists of salient object detection and salient object existence prediction task.Joint training based on multi-task learning successfully reduces the false positive rate of the model and improves the detection effect of the model in a cluttered environment.(3)A salient object detection method based on adaptive selection training is designed.The adaptive selection training algorithm based on multi-task learning uses image-level labels in order to automatically select the required saliency map to calculate the loss and train the model,this method can not only identify non-salient objects,but also reduce the interference of non-salient images on the model.