【摘要】 卷积神经网络在图像识别、目标检测等计算机视觉领域成为热门研究方向,并取得了重大进展。随着识别率的不断提高,模型深度不断加深,网络结构愈加复杂,所需的计算量和存储空间也随之大大增加,这使得卷积神经网络在资源有限的移动终端和嵌入式设备上的应用存在很多困难。因此压缩卷积神经网络,减小其占用的存储空间和计算资源成为卷积神经网络一个重要的研究方向,本文提出利用toeplitz矩阵对网络的全连接层权重参数实现共享,针对数字手写体识别网络,全连接层可学习参数压缩174倍时,模型分类准确率相较于原网络仅下降0. 74%。此外,本文提出可基于输入和输出两个角度对网络的卷积层权重参数实现循环共享,当基于输入对卷积层权重参数实现循环共享时,数字手写体识别网络的卷积层可学习参数压缩14倍,模型分类准确率仅下降0. 03%。更多还原

【Abstract】 Convolutional neural networks( CNNs) have become a popular research and made significant progress in computer vision area such as image classification,object detection and so on. With the improving of recognition precision,the corresponding CNNs are designed deeper and more complicated,which makes both computing load and storage space increase significantly. Therefore,it is hard to apply CNNs to resource-limited applications,such as mobile terminal and embedded devices. Compressing CNNs to reduce footprint storage and computing resources has become an important research. This paper provides a method to compress CNNs by sharing parameters. The toeplitz matrices are applied to the fully connected layers of CNNs and the amount of parameters reduces dramatically.Compared with that of the original CNN,the recognition rate of the proposed network to handwritten digits declines only 0. 74%,while the parameters of fully connected layers are compressed 174 X. In addition,this paper proposes to share parameters of convolutional layers in terms of input or output. When parameters shared in terms of input,the recognition rate to handwritten digits declines only 0.03% compared with that of the original network,while the parameters of convolutional layers are compressed 14 X.更多还原