Research And Implementation Of Few Shot Object Detection Algorithm In Cross-Domain Scenarios

With the vigorous development of deep learning technology and its indepth application in the field of computer vision,object detection technology has made great progress in accuracy and speed,and has been widely used in industry.General object detection technology needs to learn knowledge from a large number of labeled training samples.However,the labeling of samples often takes a lot of time and energy,and it is difficult to obtain samples in some special scenes,which makes the application cost of general object detection technology high,and the application scenarios are also limited.In recent years,few shot object detection has been proposed and researched as a solution to this problem.In the existing few shot object detection research,it is assumed that the base class samples and novel class samples come from the same domain.However,in real-world applications,they often come from different domains.Therefore,the existing few shot object detection methods generally have poor generalization performance.In order to solve the above problems,this paper designs and implements an adaptive few shot object detection algorithm,which is verified by experiments on three groups of cross domain datasets.Finally,based on this algorithm,a few shot object detection prototype system is designed and implemented.Specifically,the work and contributions of this paper are as follows:1)Through investigation and analysis,it is found that the existing few shot object detection algorithms have low generalization performance and poor application effect in cross domain scenarios.For this problem,a few shot object detection algorithm in cross domain scenarios is designed.2)Based on the few shot object detection framework meta-r-cnn,this paper introduces an image level domain discriminator after the last layer of the feature extraction network to forcibly align the global image feature representation of the source domain and the target domain,so as to reduce the domain difference.At the same time,in order to avoid the confusion of class features caused by the forced alignment of global image features,a feature filtering module based on channel attention mechanism is added to the feature extraction network to screen the features irrelevant to a specific class.3)In this paper,experimental verification is carried out on three groups of public cross domain datasets.Compared with the benchmark model,this method significantly improves the average accuracy of the model in the target domain.In the few shot cross domain scenario from Pascal VOC to clipart,the average accuracy is improved by about 8%,and in the other two experiments,the average accuracy is improved by about 3%.In addition,compared with other few shot object detection methods,it has also achieved better results.4)Based on the algorithm proposed in this paper,a few shot object detection system is developed.Users can detect novel classes by uploading a small number of novel class annotation samples.