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

As one of the important branches of computer vision,object detection aims at classifying and localize objects in a given image.Few-shot object detection,as one of the branches of object detection,aims to solve the problem that the detection performance and scene generalization of general detection model under the condition which only has a small number of samples are seriously degraded.The current challenges it faces mainly include the following three aspects:1)The inefficiency of few-shot object detection algorithm trained based on transfer learning:Currently,the mainstream backbone network weight of few-shot object detection algorithm comes from the pre-training results on large-scale supervised datasets.In most scenarios,Due to the condition that annotated data is always expensive and difficult to obtain.The initialization point of the existing few-shot object detection algorithm is not robust.2)The inefficiency of the transfer learning procedure from the base class knowledge to the novel class knowledge:At present,the mainstream methods based on transfer learnings are using weight freezing strategy to freeze the shallow layers of the network and only train the deep layers.However,these methods do not explore the general knowledge space between base classes and novel classes,resulting in low generalization of the new scenario and a serious decline in the performance of base classes.3)Currently,there still lacks the training platform for few-shot object detection algorithms:The practical application scenario lacks a complete set of platform from data collection to model construction,so that many excellent algorithms in the academic circle cannot be easily implemented and deplooyment.Based on the above problems and challenges,the main research contents of this paper are as follows:1)Designed and implemented a few-shot object detection method based on self-supervised learning.By designing a multi-instance feature contrastive learning module and fine-grained feature contrastive learning module,a better backbone network is trained,which provides a better model initialization point for few-shot object detection tasks;2)Designed and implemented a few-shot object detection method based on meta feature encoding.a meta-feature encoder is proposed to extract meta-knowledge in the training process of base classes and weight it to the final classification branch in new classes by meta-feature retrival,so as to improve the model knowledge transfer ability and model generalization.3)Designed and implemented a platform for few-shot object detection algorithms.This paper encapsulates the development procedure of few-shot object detection algorithms and provides a clear and standardized engineering construction procedure.Users can complete the construction and verification of few-shot object detection models through UI interaction.