We developed a modulation code using a complex amplitude and established a method to demodulate the code based on a convolutional neural network (CNN) for holographic data storage. The developed 20:9 modulation code consists of nine symbols, each of which contains 4 bits of data representing the symbol position on which the complex amplitude is superimposed and 16 bits of data representing the actual complex amplitude value. By solving an optimization problem, the complex amplitude signal combines four amplitude values and a different phase value for each amplitude; thus, the data are robust against amplitude and phase noise, and the amplitude and phase values are distributed over a uniform distance in the constellation diagram. Modulation tables were also optimized using a genetic algorithm. Because the occurrence of bit errors due to amplitude and phase noise must be considered when reproducing data, two CNNs separately demodulate the symbol position signal and the complex amplitude signal superimposed thereon. By inputting reproduced data and label information indicating the demodulation target, we created a compact CNN. We confirmed that the CNN demodulation can accurately demodulate both signals; moreover, the total bit errors were reduced to less than half of those for the conventional hard decision demodulation method.

我们开发了一种使用复振幅的调制代码，并建立了一种基于用于全息数据存储的卷积神经网络 (CNN) 来解调代码的方法。开发的20:9调制码由9个符号组成，每个符号包含4位表示叠加复幅值符号位置的数据和16位表示实际复幅值的数据。通过求解一个优化问题，复振幅信号组合了四个振幅值和每个振幅的不同相位值；因此，数据对于幅度和相位噪声具有鲁棒性，并且幅度和相位值分布在星座图中的均匀距离上。调制表也使用遗传算法进行了优化。由于在再现数据时必须考虑幅度和相位噪声引起的误码的发生，两个CNN分别对符号位置信号和叠加在其上的复幅度信号进行解调。通过输入表示解调目标的再现数据和标签信息，我们创建了一个紧凑的 CNN。我们确认CNN解调可以准确地解调这两种信号；此外，总误码减少到传统硬判决解调方法的一半以下。我们确认CNN解调可以准确地解调这两种信号；此外，总误码减少到传统硬判决解调方法的一半以下。我们确认CNN解调可以准确地解调这两种信号；此外，总误码减少到传统硬判决解调方法的一半以下。