Research And Implementation Of Object Detection System Based On Few-shot Learning

Object detection is one of the core tasks of computer vision.Although the algorithms in the field of object detection have made remarkable achievements in recent years,these methods cannot be directly applied to few shot scenes,which seriously restricts the application and promotion of existing algorithms.Few-shot object detection researches how to use only a few examples to classify and locate objects of unseen categories from complex background,and the challenge it faces can be summarized into three aspects:1)Currently,few-shot object detection algorithms can’t make full use of information of few-shot examples,and may have overfitting problems,so that did not achieve the optimal performance in new categories;2)The few-shot object detection algorithm usually only focuses on the generalization of the new categories,but ignores the knowledge retention of the model in the basic categories,resulting in a significant decline in the performance of the new category after transfer learning;3)Most artificial intelligence platforms are based on traditional supervised learning,but annotation data are expensive and difficult to obtain in most scenarios.At present,there are few platforms that provide services related to few-shot object detection and support the few-shot learning process.In view of the above problems and challenges,this paper mainly studies the following contents:1)Proposed and implemented a few-shot object detection algorithm based on feature enhancement and metric learning.Multi-scale feature enhancement network is designed to fuse the local and contextual features of objects,so as to obtain rich information of few-shot objects.Cosine classifier is introduced based on the idea of metric learning,and task-level feature discriminator is used to further achieve differentiated feature learning,so that improving the detection effect of few-shot new categories.2)Proposed and implemented an incremental few-shot learning scheme of knowledge decoupling,including a simple transfer learning paradigm and a plug and play dual detection head.The knowledge decoupling is realized by the parallel dual detection head,and the learning of different tasks is separated.The base class knowledge of the model is retained,and the incremental learning of few-shot object detection is satisfied.3)Designed and implemented a development platform of object detection algorithm for few-shot scenarios.It encapsulates the development process of few-shot object detection algorithm and provides a clear and standardized engineering construction pipeline.The platform provides developers with data management,model management,few-shot learning and other services to alleviate the problem of poor mobility of models in the new scenario with insufficient data.