Anchor free image target detection based on scale independent loss optimization

DOI:	10.11809/bqzbgcxb2023.06.025
Keywords:	anchor free; regression loss function; arbitrary object oriented detection; scale independence; image target detection
Abstract:	Target detection in remote sensing images is a basic and challenging problem in the field of image analysis, especially the detection of maritime targets for arbitrary objects which has received extensive attention in recent years. Aiming at the sensitivity of target detection parameters caused by the irregular scale and shape of image targets of arbitrary objects, and the false negativity in the process of non maximum suppression of the complex background and dense arrangement, this paper proposes an anchor free image target detection algorithm based on scale independent loss optimization. This paper introduces the rotating bounding box regression into a typical single stage target detector of anchor free network full convolution, and proposes the scale independent GIoU (SGIoU) loss for the bounding box regression of image target detection, which can quickly adjust the shape of the prediction box to the shape similar to the real one at the beginning. It also solves the incompatibility between the regression loss function and the final optimal target of the detector, and accelerates the convergence speed in the regression process. The data set experiment shows that the average area (mAP) of this method under various P R curves is superior to that of the existing generalized intersection/union ratio (GIoU) loss function and complete intersection/union ratio (CIoU) loss function.
Published:	2023-06-28
PDF HTML