Research On Image Salient Object Detection Based On Weak Supervision

The research of salient object detection is to detect the most interesting foreground areas in the image,and a lot of research work has been carried out in the field of computer vision.The algorithms of salient object detection are often used as a preprocessing method to improve the performance of high-level image processing tasks,such as image semantic segmentation and image caption.Therefore,it is of great significance to study the direction of salient object detection in depth.With the wide application of deep convolutional neural networks(DCNN)in the field of computer vision,the full-supervised salient object detection algorithm based on deep networks has been continuously proposed,which greatly improves the detection performance that traditional methods cannot achieve.However,training a full-supervised salient object detection algorithm based on DCNN requires a large amount of pixel-level annotated data,which requires a lot of time and labor costs.In order to alleviate the demand for large-scale pixel-level annotation data,this paper explores the salient object detection model based on weak supervision,which makes the final saliency detection performance of the model close to the detection performance of the fully supervised training model.Weakly labeled samples are less expensive and more readily available than pixel-level annotated data used by full-supervised model,such as image labels or bounding boxes.Therefore,this paper studies the salient object detection model based on weak supervision:First,the image-guided weakly supervised image salient object detection model:assuming that the dataset with only image labels,we first use the unsupervised saliency maps obtained by a traditional method,the heatmaps with location information from the image classification branch and the saliency maps predicted by the saliency detection branch to construct a dataset with pseudo-labels,and then train the proposed weakly-supervised salient object detection model,which involves the iteration of the model and the update of the pseudo-label dataset.Therefore,we can obtain the final weak-supervised salient object detection model.Second,the bounding-box-guided weakly supervised image salient object detection model: assuming that the dataset with only the bounding boxes,we first use the bounding boxes containing the location information,conditional random fieldswhich is a method of object edge refinement and the saliency maps predicted by the saliency detection branch to continuously update the pseudo-label dataset required for each round of training,then converges continuously through the model iteratively training,and finally the weakly supervised salient object detection model is obtained.Experimental results show that the two weakly supervised salient object detection models proposed in this paper are superior to the traditional methods in visualization and saliency detection evaluation metrics,and the performance is close to the full-supervised salient object detection algorithm.