关键词: 图例检测, 孪生网络, 数据增强, 单样本学习

Abstract: In view of the problems such as the difficulty of training, the slow detection and the difficulty of obtaining the training data in the existing deep learning methods, a new solution is proposed for the single sample learning problem. Based on the structure of convolutional neural network, combined with the characteristics of fixed aspect ratio and independence of legend, a new SiameseSSD detection frame is used for target detection. The framework includes a siamese subnet for feature extraction and an improved SSD subnet for classification and regression. At the same time, we use the data enhancement technology to expand the sample, then make the data set and train the model and adjust network structure and detection method to detect large-resolution construction drawings. The experimental results of this method on the construction drawing data set show that this method is a new method to solve the single sample learning task, with an accuracy of 91.3%, the detection speed reached 61 fps. Compared with the existing top level, it has certain advantages and meets the actual work needs.

Key words: legend detection, siamese network, data enhancement, one-shot learning