基于深度学习的计算机设计场景文本与非文本图片分类算法研究

Research on Text and Non-text Image Classification Algorithm of Computer Design Scene Based on Deep Learning

基于深度学习特征编码及计算机设计场景图片来进行文本图片分类。算法首先对输入的图片使用最大值稳定区域进行文字候选区域提取,然后将这些提取出来的区域输入到多类卷积神经网络中,然后将CNN最后一层的输出作为每个区域的特征;再使用词袋模型把区域特征表示成为编码,最后利用这些编码输入到支持向量机中并作最后的判定。本文算法目的是提取文字候选区域来作为算法的感兴趣区域,结合深度学习算法使得文字图片的识别更加准确,使系统具有更好的鲁棒性。在对卷积神经网络进行算法训练的时候,可以对感兴趣的区域进行聚类,最终使得原来的两种分类变成了多种分类,进而使得文字区域特征的划分更具细粒度。

This paper classifies text pictures based on deep learning feature coding and computer design scene pictures.The algorithm first extracts text candidate regions from the input images using the maximum stable region,and then inputs these extracted regions into multi-class convolutional neural networks,and then takes the output of the last CNN layer as the feature of each region.Then the word bag model is used to represent the regional features into codes,which are input into support vector machine for final determination.The purpose of the algorithm in this paper is to extract candidate areas of text as the region of interest of the algorithm,so as to have a good representation of the pictures of computer design scenes.The combination of deep learning algorithm makes the recognition of text and pictures more accurate and makes the system more robust.During the algorithm training of the convolutional neural network,the regions of interest can be clustered,which eventually turns the original two classifications into multiple classifications,thus making the classification of text region features more fine-grained.