Research On Weakly Supervised Salient Object Detection Algorithm Based On Bounding Box

Image salient object detection is an important research branch in the field of computervision.Its purpose is to focus on the target area of most interest to the human eye,to intelligently understand and calculate the main target,so as to achieve rapid analysis of the image and serve as an important prior to subsequent processing.Salient object detection is generally used as an image preprocessing technology in theoretical research,and can provide a basis for visual interaction design in practical applications.In recent years,it has received extensive attention and deep research.Existing salient object detection research is mostly fully supervised algorithms based on deep learning,which rely on truth-labeled samples and deep neural network representation learning ability to build detection models,and have good experimental performance.The labeling of truth samples requires a lot of time and costs,and the generalization in actual scenes is relatively poor,so it is also important to use weakly supervised labels(foreground target categories,text descriptions,or bounding boxes,etc.)to train deep networks.This paper mainly studies the weakly supervised salient object detection algorithm,and gives two deep learning network models that use bounding box labels for supervision.The first model is based on the calculation of neighborhood affinity.It mainly uses the affinity network to capture the similarity of saliency between local neighborhood pixels,so that the object edge pixels have a better saliency division.First,the training label of the similarity network is generated by preprocessing from the bounding box,and the affinity weight is used to calculate the loss function for supervision,and then the preprocessing result is randomly walked with the affinity matrix generated by the training to obtain the pseudo label of the salient map.Finally,the pseudo labels with higher accuracy are used to train a new salient object detection network.The second model is based on multi-task iterative optimization.This model uses a multi-task learning mechanism,which uses edge detection tasks to assist salient object detection,and optimizes some parameters of the main network through a staged network training strategy.This algorithm separately trains two network branches and then performs fusion processing,and chooses a joint loss function based on image position information and color,and implements end-to-end training through multiple loss functions.Finally,iterative optimization mechanism is used to filter the predicted saliency map to generate supervised samples for subsequent training.The above two algorithms are tested on multiple standard data sets,and compared with the advanced algorithms in the field,and satisfactory results are obtained,thus confirming the effectiveness of the algorithm in this paper.