面向小样本目标检测任务的自适应特征重构算法

Adaptive feature reconstruction algorithm for few-shot object detection

针对目标检测任务中样本量不足时新类别检测性能变差的问题,提出面向小样本目标检测任务的自适应特征重构算法。该算法包含两个模块:基础类别特征偏移缓解模块,用于获取预训练阶段基础类别的特征方向;场景特征自适应约束模块,用于根据场景特征与各类别原型特征的相关性确定当前场景对于某些类别的偏好,从而自适应地调整基础类别偏移方向对实例特征的影响。试验结果表明,在PASCAL VOC和MS COCO数据集上,该模型对于小样本目标检测任务的检测能力均优于对比算法,在保证对于基础类别实例检测能力的基础上,对新类别的检测精度最高可分别提升12.4%与2.1%。本研究提出的模型可以保证对于基础类别相关实例的检测能力,并提升新类别实例检测性能。

Aiming at the problem that the detection accuracy of novel categories decreased when the sample size was insufficient in object detection tasks,an adaptive feature reconstruction algorithm was proposed.The algorithm contained two modules,one of which was the base categories′feature offset′s alleviation module,which obtained the offset direction of the base categories in the pre-training stage.The other was the scene feature adaptive constraint module,which learned the correlation between scene features and the prototype features of each category to determine the preference of the current scene for certain categories,and adaptively adjusted the influence of base categories offset direction on instance features.Experimental results showed that the performance of the proposed algorithm was superior to the comparison algorithms of few-shot object detection tasks on PASCAL VOC and MS COCO datasets.On the basis of ensuring the detection capability of the base categories′instances,the detection accuracy of novel categories′instances could be improved by 12.4%and 2.1%,respectively.The proposed model could guarantee the detection ability of instances of base categories and improve the detection performance of instances of novel categories.