基于改进Faster R-CNN的小样本目标检测算法OA

In order to solve the problems of limited information that can be learned by the model due to the scarcity of training samples in few-shot scenarios,the performance of the model is restricted by the scale change of the object,and the catastrophic forgetting of the base class knowledge,a few-shot object detection algorithm with improved Faster R-CNN was proposed by following the transfer learning method.Firstly,in the feature extraction stage,the central coordinate attention mechanism was added to guide the model to learn the core features of the object,and multi-scale fusion and deformable convolution were introduced to enable the model to extract rich feature information on the feature map at different scales.Then,a gradual RPN was introduced,and a two-stage regression method was used to gradually optimize the generated regional suggestions.Finally,the knowledge compensation module was introduced to decouple the prior knowledge from the new knowledge with the help of the decoupling mechanism,so as to maintain a good memory of the base class.Experimental results show that,in the three novel class splits of the PASCAL VOC,compared with the mainstream small sample object detection method FSODCR,the proposed algorithm achieves an average increased in nAP by 8.37 percentage points and 6.63 percentage points in the 2-Shot and 5-Shot settings respectively;on the MS COCO,compared with FSODCR,the proposed algorithm achieves an average increase in nAP and nAP75 of 1.2 percentage points and 0.9 percentage points respectively under the 10-Shot,and nAP and nAP75 have increased by 1.2 percentage points and 0.8 percentage points under the 30-Shot,which significantly improves the accuracy and stability of object detection under few-shot conditions.