基于增量学习的目标检测研究

目标检测技术已经被广泛应用于行人检测、人脸识别等诸多领域。随着社会生活与工业发展中新需求的出现，目标检测的对象与要求也随之出现新的变化。若直接用旧模型训练新类别可能会导致灾难性遗忘缺陷。因此，增量目标检测逐渐成为一个热门的研究方向。总结了常用的数据集和模型评价指标，研究了增量目标检测技术，将增量目标检测分为基于知识蒸馏的目标检测模型、基于回放的增量目标检测模型、开放世界目标检测模型，指出目前增量目标检测存在新增可识别类别的数据有限、新的可识别类别增加会导致模型准确率降低、小目标检测困难、检测速度慢等问题。通过比较最新增量目标检测模型，提出未来增量目标检测应从优化知识蒸馏方式、加强旧目标类别范例样本选择、更好地结合Transformer网络等几个方面进行改进。

Research on Object Detection Based on Incremental Learning

Object detection technology has been widely applied in many fields such as pedestrian detection and facial recognition. With the emergence of new demands in social life and industrial development, the objects and requirements of object detection have also undergone new changes. Training new types directly with old models may lead to catastrophic forgetting defects. Therefore, incremental object detection has gradually become a popular research direction. This research summarized commonly used datasets and model evaluation indicators, studied incremental object detection techniques, divided incremental object detection into knowledge distillation based object detection models, replayed based incremental object detection models, and opening world object detection models. It is pointed out that there are currently limited data for newly added target categories in incremental target detection, an increase in new identifiable categories that can lead to reduced model accuracy, more difficulty in detecting small targets, and slow speed in detection. By comparing with the latest incremental target detection models, it is proposed that future incremental target detection should be improved by optimizing knowledge distillation methods, strengthening the selection of old target category example samples, and better combining Transformer networks.