Object Detection Methods Based On Few-shot Learning

Deep learning models have achieved state-of-the-art performance on object detection task.However,deep learning models rely heavily on a large number of labeled data.In real life,manual data tagging is time-consuming,laborious,expensive.And in some application areas,there is already insufficient data accumulation.Inspired by human fast learning ability,few-shot learning under the branch of deep learning hopes that after the model has learned a large amount of labeled data of the base classes,for the novel classes,it only needs a small number of labeled samples to quickly learn.Therefore,this paper intends to adopt few-shot learning to propose an efficient and robust object detection method.The specific tasks are as follows:(1)The few-shot object detection model based on fine-tuning first uses a large amount of labeled base classes data to train a detector.Then freeze the feature extractor in the model,and fine-tune the classifier and regressor on the novel classes with K shots per class.Due to the small number of labeled samples in the fine-tuning stage,it is easy to cause over-fitting problems.In response to the above problems,on the basis of the model proposed by Wang et al.,we proposes a feature enhancement module based on K-combined mean.The specific method is to use the feature extractor trained in the first stage to obtain the features of K shots.Take these K features each C _K~m(1?m?K) into a group,Calculate the corresponding average characteristics for each group,and add it to the feature set of each class.In the fine-tuning stage,the enhanced feature set is used to fine-tune the classifier.The experimental results show that the method is effective.(2)The input of the few-shot object detection model based on metric learning is the query-target image pair,and the output is the regions in the target image that is similar to the query image.When the model is trained on the base classes,the labels of the query image and the target image are known,and the purpose of training is to learn the similarity metric between the query image and the target image.It is class-agnostic,and then directly apply the learned metrics to the novel classes during the test.Because the target image has a rich background and contains multiple classes of foreground objects,the algorithm is prone to false detection results for the classes that have correlations.In response to this problem,this paper proposes a data enhancement based on object exchange.The specific method is to exchange the objects in the target image and the query image that belong to the same class when training on the base classes,and generate a new query-target image pair as input.The experimental results prove that this method provides a more diverse comparison sample for the learning of the measurement module,and improves the detection accuracy on the new classes and the base classes.(3)Because the few-shot object detection algorithm based on metric learning only pays attention to the similar area between the query image and the target image,this type of algorithm has poor detection accuracy for different classes of objects with similar appearance and objects with large appearance differences but belonging to the same class In response to this problem,this paper proposes a knowledge transfer algorithm based on category names.The specific method is to use the category name corresponding to the query image as semantic knowledge,calculate the corresponding word vector as the feature of the semantic space,and merge it with the image feature of the query image.The experimental results prove that this method increases the dissimilarity between classes and the similarity within classes by using the category name,and improves the detection accuracy on the base classes and the novel classes.