基于多元特征异构集成深度学习的图像识别模型及其应用  

Image Recognition Model Based on Multivariate Feature Heterogeneous Ensemble Deep Learning With Its Application

作　　者：汤健 田昊 夏恒 王子轩 徐喆 韩红桂 TANG Jian;TIAN Hao;XIA Heng;WANG Zixuan;XU Zhe;HAN Honggui(Faculty of Information Technology,Beijing University of Technology,Beijing 100124,China;Beijing Laboratory of Smart Environmental Protection,Beijing 100124,China)

机构地区：[1]北京工业大学信息学部,北京100124 [2]智慧环保北京实验室,北京100124

出　　处：《北京工业大学学报》2024年第1期27-37,共11页Journal of Beijing University of Technology

基　　金：国家重点研发计划资助项目(2018YFC1900800-5)。

摘　　要：随着城市矿产资源循环利用技术的不断发展,废旧手机回收已成为当前研究热点。受限于计算资源和数据资源的相对缺乏,目前基于线下智能回收装备的废旧手机识别精度难以达到实际应用。针对上述问题,提出一种基于多元特征异构集成深度学习的图像识别模型。首先,利用字符级文本检测算法(character region awareness for text detection,CRAFT)提取手机背部字符区域,再利用ImageNet预训练的VGG19模型作为图像特征嵌入模型,利用迁移学习理念提取待回收手机的局部字符特征和全局图像特征;然后,利用局部特征构建神经网络模式光学字符识别(optical character recognition,OCR)模型,利用全局和局部特征构建非神经网络模式深度森林分类(deep forest classification,DFC)模型;最后,将异构OCR和DFC识别模型输出的结果与向量组合后输入Softmax进行集成,基于权重向量得分最大准则获取最终识别结果。基于废旧手机回收装备的真实图像验证了所提方法的有效性。With the continuous development of recycling technology for urban mineral resources,recycling of used mobile phones has become a hotspot for current research.Restricted by the relative lack of computing resources and data resources,the accuracy of used mobile phone recognition based on current off-line intelligent recycling equipment is difficult to meet the practical application.Therefore,an image identification method based on multivariate feature heterogeneous ensemble deep learning method was proposed.First,the character region on the back of the mobile phone was extracted by using character region awareness for text detection(CRAFT) algorithm,the VGG19 model pre-trained by ImageNet was used as the image feature embedding model,and the local character feature and global image feature were extracted by using the transfer learning mechanism.Then,the optical character recognition(OCR) character recognition model based on neural network(NN) mode was constructed by using the local feature,and the improved deep forest classification(DFC) model based on non-NN model was constructed by using the global and local features.Finally,the outputs of heterogeneous OCR and the DFC model were integrated and fed into the Softmax to ensemble,and the final recognition result was obtained based on the criterion of maximum category weight vector.The effectiveness of the proposed method was verified based on real images of used mobile phone from recycling equipment.

关 键 词：废旧手机 图像识别 迁移学习 多元特征 OCR 深度森林 异构集成 

分 类 号：U461[机械工程—车辆工程] TP308[交通运输工程—载运工具运用工程]