基于递进卷积神经网络的台标识别及其并行化  

Channel logo recognition based on progressive convolutional neural networks and its spark implementation

作　　者：许子立 姚剑敏[1] 郭太良[1] 

机构地区：[1]福州大学平板显示技术国家地方联合工程实验室,福建福州350002

出　　处：《电视技术》2016年第5期67-73,共7页Video Engineering

基　　金：国家"863"重大科技专项(2013AA030601);福建省科技重大专项(2014HZ0003-1)

摘　　要：针对台标的视觉特征,提出一种基于递进卷积神经网络的台标识别算法。该网络不仅有对图像特征进行隐性提取的卷积层和采样层,还包括识别常规台标的泛化模块和识别偏差台标的特异模块。针对串行卷积神经网络训练耗时长的缺点,提出基于Spark的并行递进卷积神经网络算法,采用数据共享及批处理方式对算法模型进行并行化处理。实验证明,递进卷积神经网络算法对台标进行识别能达到98%的正确率,多节点并行化卷积神经网络相比于单节点模型能有效缩短80%以上训练所需的时间。Referring to the character of channel logo, a channel logo recognition based on progressive convolutional neural networks is proposed. The network is not only have convolution layer and sampling layer to implicit extract image features but also includes the generalization module and subject-specific modules. Because of problems of time-consuming is too large for training serial convolution neural network, a parallel progressive convolutional neural networks algorithm based on Spark is proposed, which uses data sharing and batch mode. Experiments show that the performance of the channel logo recognition has been well improved, the method achieves 98% correct recognition rate, multi-node parallelism convolution neural network can shorten more than 80% training time.

关 键 词：台标识别 卷积神经网络 递进 并行 SPARK 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]