Research On The Target Contour Detection Algorithm Based On The Number Of Fluctuations

Object contour detection is one of the most foundational,significant and challenging problems in the field of computer vision.Although deep learning and other technology continue to research deeply in various directions,most of them are still dedicated to edge detection,and compared with the research of object detection,semantic segmentation and instance segmentation,object contour detection is less explored.In terms of data,the current state-of-the-art algorithms also need to be trained through a large amount of supervised data,and most tasks require specific label data,such as contour detection data set.With the explosive growth of data,and the cost of manual labeling increases,and the correctness cannot be guaranteed.In view of the above two problems in contour detection,this paper first proposes to use weakly supervised learning technology to process the contour detection data set.The idea is to combine arbitrary bounding boxes and edge detection algorithms to obtain weak edges,and then suppress the edges of the interior and background of them to obtain weakly annotated contours,which can reduce manual consumption and expand the data range.However,a data set composed entirely of weakly annotated contours will inevitably affect the training and detection performance of the model.Although this paper proposes two limiting methods for bounding boxes selection before obtaining weakly annotated contours,the experimental results show that the performance of various contour detection under the weakly annotated data obtained by this method can only approach the performance under the dedicated data set.For this reason,this paper introduces fluctuation times and gives it a new concept,and proposes two methods of repairing data in the process:filling and re-defining,which adds useful information of weak annotation contours.Finally,this paper improves the single-shot instance segmentation(Polar Mask)model represented by polar coordinates,and proposes a modeling method based on fluctuation times in polar coordinates.Experimental results prove that some results of this method are better than known algorithms.