基于矢量量化压缩密集连接层的图像压缩研究

针对卷积神经网络(CNN)在图像压缩耗费较大存储空间问题,文中通过研究压缩CNN参数的矢量量化方法解决了CNN模型的存储问题。通过压缩密集连接层的存储方式使得矢量量化方法比现有的矩阵分解方法更具优势。将k-均值聚类(KM)应用于权重和乘积量化可以在模型大小和识别精度之间取得较好的权衡。实验结果表明,结构化量化方法的效果明显优于其他方法,通过对图像压缩检索验证了压缩模型的泛化能力。

Image compression based on vector quantization compression dense connection layer

To solve the problem that convolutional neural network(CNN)consumes much storage space in image compression,the storage problem of CNN model is solved by studying the vector quantization method of compressed CNN parameters.By compressing the storage mode of dense connection layer,the vector quantization method has more advantages than the existing matrix decomposition method.Applying k-means clustering(KM)to weight and product quantization can achieve a better trade-off between model size and recognition accuracy.The experimental results show that the effect of structured quantization methods is better than other methods,and the generalization ability of the compression model is verified by image compression and retrieval.