Weakly supervised object detection algorithm based on feature enhancement and loss optimization

DOI:	10.11809/bqzbgcxb2023.06.028
Keywords:	weakly supervised object detection; multi instance learning; efficient selective channel attention; regression loss; dynamic optimization
Abstract:	Aiming at the problem that only the most discriminative part of an image can be detected in weakly supervised object detection and the training process is easy to fall into the local optimum, this paper proposes a weakly supervised object detection algorithm based on feature enhancement and loss optimization. In this algorithm, an efficient selective channel attention module is proposed to expand the most discriminative sample object area by improving the interaction ability of local information through the selection of correlation channels. In addition, by applying specific dynamic weights to the network regression loss function, it can automatically weaken the influence of the inaccuracy of the pseudo labeled bounding box in the regression branch and improve the accuracy of object location. The experiments on PASCAL VOC 2007 and PASCAL VOC 2012 datasets show that, compared with other similar algorithms, the proposed algorithm can effectively improve the accuracy of the weakly supervised object detection. At the same time, because the extra training parameters and computational burden introduced by this algorithm are almost negligible, it also has good efficiency.
Published:	2023-06-28
PDF HTML