Research On One-shot Object Detection Based On Convolutional Neural Network

As a branch of computer vision,object detection plays a great role in various fields such as automatic driving,monitoring and security.At present,the deep learning methods used in the field of object detection need the support of a large amount of data.However,a large amount of annotation data can not be obtained in many scenes,such as the detection of special signs.One-shot object detection task means that only one sample is known for the specific class.For this task,this thesis researches metric learning and cross-domain learning:1.Based on the metric learning method,a one-shot object detection algorithm based on high-order co-attention and low-rank representation is proposed in this thesis.Starting with the in-depth mining and utilization of one-shot information,firstly,the target image and query image are input into the Siamese network to extract features,and then the high-order co-attention mechanism is used to fuse features,extract and utilize the deep information of features,so as to enhance the learning ability of neural network for one-shot image.At the same time,a low-rank representation structure is designed.The low-rank matrix is used to jointly represent the candidate feature and query feature,enhance the extraction and classification process of candidate feature,and effectively improve the object detection ability of neural network for one-shot.Experiments on public datasets show that the improved algorithm proposed in this thesis exceeds the best one-shot object detection algorithm,and verifies the effectiveness of high-order co-attention and low-rank representation algorithm.2.Based on the cross-domain learning method,a cross-domain one-shot object detection algorithm based on image scale and instance scale is proposed in this thesis.The source domain dataset is added to enhance the one-shot data,and the cross-domain classifier is used to reduce the domain offset between different domain datasets.The specific approach is to construct the domain classifier of the target image and query image,enhance the domain adaptability of the network to the background and target at the image scale,and design the domain classifier of the candidate box instance scale to enhance the robustness of the candidate network to different classes.Experiments in three cross-domain scenarios show that the proposed domain classifier can effectively enhance the domain adaptability of the one-shot object detection network.