Research On Few-shot Object Detection Based On Deep Learning

In recent years,deep learning has laid a solid foundation for the rapid development of computer vision algorithms,and object detection algorithms based on deep neural networks have gradually been applied to all aspects of production and living.However,with the development of technology,the research of object detection algorithms has gradually shifted from scenarios with sufficient labeled data and strong supervision to scenarios with few samples with long tail distribution of categories.Therefore,how to use a small amount of data to achieve high-precision object detection has become a hotspot in both domestic and foreign research field.However,without enough training samples,deep neural networks are prone to problems such as difficulty in convergence,overfitting,and poor generalization performance.And because the number of samples corresponding to certain categories is very small,the change in the scale of those samples will have innegligible impact on the detection accuracy.When the distribution of sample numbers between different categories is extremely unbalanced,categories with few data are easily overwhelmed by negative gradients generated by other categories during training procedure,resulting in lower detection accuracy.To address above issues,this thesis proposes and implements few-shot object detection algorithm based on meta learning and transfer learning.Firstly,this thesis presents a few-shot object detection algorithm based on meta learning.The algorithm uses Center Net,which is an one-stage anchor-free detection model,as based detection model,and reconstructs the object detection paradigm based on meta-learning in the few-shot scenarios.Specifically,the algorithm proposes scale insensitive feature fusion module for the problem of unbalanced distribution of sample scales in the few-shot scenarios,which allows the model to more accurately extract features of different scales of samples,and significantly improving the detection accuracy on novel classes.In addition,in order to more effectively extract and fuse the meta-knowledge in the support set,mutual global context operation module is proposed,which could extract the global semantic information of the support set to improve the detection accuracy.Finally,Pixel Loss is also proposed to enhance the small object detection ability of the model.Extensive experiments implemented on high-performance servers and edge computing devices demonstrate that the proposed algorithm in this thesis not only outperforms than mainstream algorithms based on metalearning on accuracy,but also has faster detection speed and lower memory occupation on edge computing devices.Next,this thesis proposes a few-shot object detection algorithm based on transfer learning.This algorithm has conducted further research on the unbalanced distribution of the number of region proposals in the two-stage object detection model under the few-shot scenarios,and proposed corresponding strategies for different phases of network training.In the base training phase,the parameters of novel classes is susceptible to the large number of negative gradients generated by the base classes during the training procedure,which causes a low detection accuracy on novel classes.This algorithm alleviates the impact of the negative gradient on novel classes by improving the loss function of the classification module,thereby improving the detection accuracy of novel classes.In the fine-tuning phase,by comparing the number of region proposals of different classes,we found that the number of region proposals of base classes greatly exceeds that of novel classes during the training procedure,which makes it difficult to improve the detection accuracy of the novel classes.In order to increase the number of region proposals belongings to novel classes,the RPN(Region Proposal Networks)structure in the two-stage target detection model is redesigned.Specifically,the Refinement Branch is proposed to help the network filter out more region proposals belonging to novel classes,which could significantly improves the detection accuracy of novel classes.Extensive experiments demonstrate that proposed algorithms in this thesis can increase the number of region proposals belongings to novel classes,and outperforms the latest few-shot object detection Fs Det by roughly 1%?6% on the accuracy of novel classes under the PASCAL VOC benchmarks.The proposed few-shot object detection algorithm studied and implemented can be applied to intelligent industrial defect detection task,and it also has important theoretical significance and practical value for the long tail and few-shot detection scenarios in computer vision applications.