Compared to traditional noise reduction methods, deep learning-based denoising algorithms can denoise unseen types of noise, enhancing the visual quality of denoised images. However, deep learning networks require a large number of images for training, which is often not available in biomedical experiments where there is a lack of samples to train neural networks. In response to this situation, a zero-shot learning convolutional neural network is proposed, which generates different images through subsampling technology to be used alongside noisy images for the residual network to learn from. The loss function in the residual network combines residual loss, regularization loss, and guidance loss, while attention mechanisms are also incorporated into the convolutional layers. In the case of noise standard deviations of 10,25 and 50, the PSNR of the algorithm is 35.43dB, 29.93dB and 24.81dB. The experimental comparison with other noise reduction algorithms shows that this algorithm has better noise reduction effect in the case of low noise, and also shows comparable performance with other models in the case of high noise, which proves its high efficiency and stability in the case of limited samples.