A Few-shot Object Detection Algorithm Based On Data Augmentation And Meta-learning

As one of the highly researched directions in the field of computer vision,object detection has widely applied in various domains.Nevertheless,in some real-world scenarios,data sets are often small,making it difficult to train highly accurate target detection models.Data augmentation is an effective method to solve the problem of data scarcity by transforming the training data to generate more training samples in order to improve the generalisation ability of the model.Meta-learning can be employed in object detection to quickly train highly accurate models from small data samples.Thus,to address the challenges of limited samples,poor model generalization,and low accuracy in realistic scenarios,this thesis presents a few-shot object detection algorithm based on data augmentation and meta-learning.This algorithm generates diverse training samples through data augmentation and employs meta-learning to rapidly learn new tasks,achieving fast and accurate object detection tasks with small data sets.Key contributions of this work include:Firstly,this thesis proposes an improved data augmentation method called saliency CutOut.The method combines a saliency detection algorithm on top of CutOut to generate new samples by selective masking of images.Not only does it increase the diversity of the data,but also avoids losing key features of the data.The paper then compares it with traditional data enhancement methods such as random flip,random crop,random brightness contrast and the CutOut method.Secondly,this thesis improves on the Meta R-CNN based on meta-learning and proposes a feature extraction network based on the convolutional block attention module.In this thesis,the convolutional block attention module is added to the residual neural network so that the model learns the importance distribution of the feature map and reduces the attention to useless information.In addition,this paper introduces feature pyramids into the neck network of the region suggestion network.This approach plugs the area suggestion network into the feature pyramid hierarchy to realise the feature pyramid structure and improve the detection accuracy in small-sample scenarios.Thirdly,this thesis implements a web-based system platform that combines the algorithm model with front-end interaction techniques to provide a highly user-friendly and easy-to-operate interface for users to easily upload images and obtain detection results.The user can also train the model in the visual interface,providing more options for the user.The experiments show that saliency CutOut improves somewhat on each category,with an average accuracy improvement of 2 ～ 5%.Although different data enhancement strategies can all improve the detection accuracy of the object category to some extent,the saliency CutOut has the best robustness.The algorithms in this thesis have an average0.6 ～ 5% improvement on the VOC dataset and 0.7 ～ 1.2% improvement on the COCO dataset.The object detection platform has higher user-friendliness and ease of operation.