Research Of Object Detection Based On Few-Shot Learning

Object detection is one of the classic tasks in computer vision,which has a wide range of applications in practical production and life.At present,object detection methods based on deep learning rely on a large number of labeled data,and it is difficult to obtain a large number of labeled data in some scenarios,such as: data of rare disease patients,data of endangered animals,etc.Therefore,it is of great significance to research the few-shot object detection to get rid of the dependence on large-scale labeled data.At present,few-shot object detection methods can be divided into three categories: 1)data enhancement based methods,2)meta-learning based methods,and 3)transfer learning based methods.At present,these methods have achieved good performance in few-shot object detection,but there are still the following shortcomings:1)The previous methods regard the RPN module as a class independent module,but the RPN model has implicit class correlation.Directly applying the RPN module trained on the base class to the novel class will result in the novel class proposal being used as the background,thereby reducing the recall rate of the novel class.In the ROI module,the classifier has low classification accuracy due to data scarcity.2)In the initial stage of most previous methods,the model is trained on the base dataset to detect the base class,but there is an issue of inappropriate use of data during this process.There are multiple instances of categories in a single image,while the images in the base dataset only annotate the base class instances,resulting in the model using instances of other unlabeled categories outside of the base class as background in the initial learning stage,which leads to cognitive bias in the learning process of the model.It is difficult for the model to correct this bias when facing novel classes with only a small amount of annotation data.In response to the above issues,this thesis proposes two solving algorithms:1)Few-shot object detection method based on collaborative RPN and collaborative ROI.The main contributions are as follows: First,superclass attributes are designed to as-sist in the detection of novel classes.Secondly,the collaborative RPN module is designed to fuse the foreground confidence and superclass attribute similarity of features to avoid the novel class proposal being rejected as a background and improve the recall rate of novel classes.Then,the collaborative ROI module is proposed to combine the corrected metric-based classifier and linear classifier,endowing the combined classifier with high generalization and learning ability,and improving the classification accuracy.Finally,a large number of experiments are carried out to prove the effectiveness of this method,and the functions of each module of this method are demonstrated through ablation experi-ments and visualization experiments.2)Few-shot object detection method based on the contrastive consistency mutual enhancement structure.The main contributions are as follows: First of all,a mutual en-hancement structure is designed to make full use of the unlabeled novel class instances in the base class dataset.Secondly,contrastive learning is introduced to construct the pro-posal contrastive consistency module to help the model resist the influence of noise labels and improve classification ability.Then,a novel data synthesis module is proposed to gen-erate additional training data to stabilize the training of the model.Finally,a large number of comparative experiments are carried out to verify the effectiveness of this method,and the role of each module of this method is explained through ablation experiments and visualization experiments.