基于自注意力机制的神经网络多模光纤图像传输

multi-mode fiber binary image transmission with high accuracy is realized by using multi-head self-attention based neural network. Compared with the common CNN-based Real-valued ANN or U-net methods, the method makes use of the multi-head self-attention mechanism implemented by sliding Windows to achieve a more stable image transmission with fewer parameters and a wider receptive field in the feature extraction module. We built an optical experimental platform to collect multi-mode and multi-mode optical fiber transmission data, and realized the high efficiency and high accuracy of binary image transmission, with the transmission accuracy reaching 94.32% on the 5000 size Mnist dataset. Our findings pave the way for robust multimode fiber image transmission schemes that have the potential to be referenced in a variety of highly conditional image transmission, paving the way for possible high-precision image transmission tasks, such as multimode fiber endoscopy, or other remote multimode fiber image transmission applications