一种元卷积网络和持续学习融合的洞库类目标识别方法

洞库类目标是高价值识别目标，针对洞库类目标样本数据难以获得、样本内部数据相似度较高、人工设计识别特征方法局限性较大、普通深度网络需要海量数据等问题，提出了结合元学习和深度卷积网络的元-卷积网络（MCNN），并融合持续学习理论的洞库类目标识别方法（MCNN-LLS）.首先结合深度卷积网络、元学习的理论建立元-卷积网络，该网络可利用旧知识指导新知识的训练，利用小样本数据即可训练得到识别能力较高的深度洞库模型；然后融合持续学习理论，建立持续学习系统（LLS），设计专家审核模型判别深度洞库模型的识别结果，并引入潜在任务、模型异步更新等方法，达到模型持续学习、持续更新的效果.实验表明，本文方法所需样本数量少，对洞库类目标识别准确率高，且识别能力可随识别过程中新数据的积累逐步提高. 

Cave Targets Recognition on Meta-Convolutional Networks and Lifelong Learning

Cave target is the high-value recognizing target. According to the difficulty about cave target data collection, the high data similarity, the limitation of the artificial feature, and the deep neural networks needs massive data, a method of combining meta-convolutional network and deep convolutional networks named meta-convolutional networks(MCNN), and combining lifelong learning was proposed (MCNN-LLS). Firstly, a meta-convolutional network was established by combining deep convolutional network and meta-learning. This network can use old knowledge to guide the training process, and can use the small sample to train an ideal cave detection model. Then combining lifelong learning and establishing the lifelong learning system (LLS), designing the expert review model to identify the recognition results by the cave detection model, and introducing potential tasks, model asynchronously update to reach the effect of model sustainable updating. Experiments show that this method only needs small sample, has high accuracy of recognizing cave target, and the recognition effect can gradually increase with the accumulation of new data.