Abstract: In recent years, deep convolutional networks are widely used in image recognition, speech recognition and natural language processing and other fields, which have achieved very good results. In this paper, in order to solve classification problems with all the samples being unlabeled data, deep convolutional neural network is improved and the corresponding learning algorithm is given，with the k-means clustering device replacing the classifier in deep convolutional network and adopting convolutional auto-encoder. The improved learning algorithm is used to solve the stitching of scrapped paper. The experiments show that, this method is effective and feasible, which improves the accuracy and robustness of the stitching of scrapped paper.

Key words: convolutional neural networks, k-means clustering, stitching of scrapped paper, convolutional auto-encoder, deep learning

摘要： 近年来，深度卷积网络在图像识别、语音识别和自然语言处理等领域广泛使用，取得了很好的效果。为解决全部样本均为无标签数据的分类问题，对深度卷积神经网络进行了改进，采用卷积自动编码器学习输入样本的特征，用k-均值聚类器代替深度卷积网络中的分类器，建立了改进的深度卷积网络结构，给出了相应的学习算法，将其用于解决碎纸片拼接问题。实验表明，该方法有效可行，提高了碎纸片拼接的准确性和鲁棒性。

关键词: 卷积神经网络, k-均值聚类, 碎纸片拼接, 卷积自动编码器, 深度学习