Research On Few Shot Incremental Scene Object Detection Method

Traditional object detection models need to be trained with a large amount of data under supervision,while humans can gradually learn new concepts with only a few examples.With the rapid development of the Internet of Things,lightweight devices such as edge devices and embedded devices have a small amount of computing and storage capabilities.In order to achieve resource information collaboration,providing edge devices with visual detection capabilities has become a development trend of artificial intelligence.This thesis intends to study incremental object detection in data sparse scenarios.The overall algorithm framework includes a mapping fine-tuning layer for feature decoupling and a knowledge extraction module for incremental transfer.In the feature decoupling mapping layer,according to the current fine-tuning scheme,the design is based on the base class training set,and the decoupling feature extraction process is a category-independent extraction module and a category-related extraction module,and the features are fully extracted.In addition,for the new class data sparse scene,the new class category is easy to overfit.A object adaptor is proposed to fine-tune the feature map layer adaptive category-independent feature extraction with the base class data,and fully extract the category-independent features.This method achieves better results in image representation and is more robust.In the knowledge extraction module,this thesis uses the pre-trained network copy as the teacher model to guide the backbone network to learn the similarity between the base class and the novel class,and retain the base class knowledge at the Logit level through the incremental classifier;in addition,considering the possibility of new class scenarios Containing old-class categories,we use masks to occlude new-class objects to align operations with knowledge extractors at the feature level such that they preserve the base-class distribution similarity.This method fine-tunes the new class through the incremental transfer strategy,which can effectively identify the new class and reduce the forgetting of the base class knowledge,thus solving the problem of catastrophic forgetting in the incremental stage.Experimental results show that the method in this thesis has better performance.In this thesis,by studying the learning method in the few shot scene,using feature decoupling,it can effectively extract the target features in the data sparse scene.And with the help of knowledge distillation technology,under the premise of retaining the accuracy of new class recognition,less old knowledge is forgotten.Realizing the online registration of sparse new class data for object detection in dynamic scenes has high application and research prospects.