Research Few-shot Object Detection Algorithm Via Deep Learning

As one of the basic tasks in the field of computer vision,object detection has been widely used in production and life.With the rapid development of deep learning,the detection accuracy and detection speed of object detector are constantly improving.Models based on deep learning generally require a large amount of labeled data to converge,but these training sample data are difficult to obtain in many real scenarios.In order to make the model complete the training with only a few samples,the researchers put forward the few-shot learning method.Compared with few-shot classification tasks,the application of few samples in the field of target detection is more challenging,because detection tasks need to be classified and located at the same time.In order to solve the existing problems in the field of few-shot object detection,this paper proposes a new few-shot object detection method and applies the distillation algorithm to few-shot object detection.The main research of this paper is as follows:(1)In this paper,a few-shot object detection algorithm combining classification correction and sample amplification is proposed to solve the problem that the use of amplified samples in fewshot learning tasks will cause data distribution deviation and the mutual influence between classification tasks and location tasks in object detection.Large data sets other than the base data are used to reduce the deviation between the data distribution and the real distribution,and new samples are amplified from the revised sample distribution.In gradient back propagation,limiting factors are used to control the backbone network to receive base and new class information in different degrees.Finally,a strong classification network is used to correct the classification score output by the object detector.Experimental results show that the method proposed in this chapter is effective in improving the detection performance of the few-shot object detector.(2)In this paper,a knowledge distillation method for object detection based on foreground enhancement and spatial frequency is proposed to emphasize the importance of global information in object detection and to ease the imbalance in the number of foreground and background regions.The attention mechanism is used to enhance the expression of foreground area in the feature,so that the student network also pays attention to the foreground area.Fast Fourier convolution is used to make the student network inherit the global information from the teacher network.Finally,the losses of classification tasks and regression tasks are calculated respectively,and the full connection layer is used to assign weights to all losses adaptively.This knowledge distillation algorithm is applied to few-shot object detection by a specific training method.Experimental results show that the distillation method proposed in this chapter is effective in improving the performance of the object detector and the training strategy is also effective in applying the distillation method to the few-shot object detection task.