Research On Few-Shot Object Detection Technology

Object detection has always been one of the hot topics in the field of computer vision.It is also a necessary prerequisite for a large number of advanced vision tasks,and has a large number of applications in practical tasks.Most of the current mainstream object detection algorithms rely on a large amount of label data for training.However,it is difficult to obtain a large amount of data and corresponding labels for some application scenarios,such as military and some security areas.Therefore,the object detection technology of few-shot samples has very important application value in these fields.In this paper,we propose a few-shot object detection method based on data enhancement and model improvement.The main contents of the research are as follows:(1)We propose a data enhancement method for few-shot learning.By combining the false samples generated by the GAN network with the data enhancement method,expand the small sample data under the premise of ensuring the target quality of the sample.Starting from the source of the difficulty of few-shot learning,improve the detection precision of the detection module.Through the test on NWPU VHR-10 dataset,after using our data enhancement method proposed in this chapter can improve the average precision by nearly 6%.(2)We propose a quadratic discriminant network based on siamese neural network.Referring to the working principle of sliding window,combined with the similarity calculation characteristics of siamese network,determine and adjust the output of the detection algorithm,so as to improve the detection results.Through the test on NWPU VHR-10 dataset,the correction method in this chapter can improve the average precision by nearly 2%.(3)We propose an object detection method based on attention mechanism.By adding the channel and spatial attention module,we design a new structure to replace the original residual structure in the feature extraction network,in order to solve the problem that the features of target itself and location information cannot be completely extracted in the few-shot situation.Strengthen the attention to the target itself,and weaken the influence of irrelevant information,improve the detection of few-shot target by improve the model structure.Through the validation on the public dataset Pascal VOC 2007,the mean average precision of the proposed method is 81.8%.The results show our method has the better performance than the conventional deep learning detection algorithm on few-shot data sets.The average precision of each category and mean average precision are improved to a certain extent.What's more,the train and inference time of detection model are not increased too much.