<p>Hybrid quantum-classical neural networks (HQCNN) have emerged as a frontier technology for addressing complex computational problems. However, the presence of qubit flip noise (QFN) in quantum channels has the effect of diminishing the effectiveness of quantum computing. In this paper, we present the construction of a HQCNN model incorporating a QFN, based on an analysis of the impact of noise on model performance. The investigation compares the classification accuracy, F1 score, and loss values under different noise intensities, with and without transfer learning, to ascertain the influence of QFN on the model. The experimental results demonstrate that the recognition accuracy and loss values on the MNIST dataset surpass those of the classical CNN and FC models, and the classification performance improves as the depth of the PQC increases. However, as the noise intensity increases, the model’s accuracy and F1 score gradually decline, while the loss value rises, indicating that QFN has a significant negative impact on the performance of HQCNN. Moreover, the experimental results on the Fashion-MNIST dataset also confirm the generality of this trend. Finally, transfer learning is applied to the HQCNN model. On both datasets, transfer learning consistently demonstrates enhanced robustness in medium- to high-noise environments. In low noise intensity (noise 0–0.2), smaller performance gaps in the models and more limited interference of noise with the models. In moderate noise intensity (noise 0.4–0.6), the benefits of transfer learning are emerging to mitigate the disruption caused by noise. In high-noise environments (noise 0.8–1.0), the interference of quantum noise is very obvious, and since migration learning uses model parameters to train the source task initialization, it makes model training smoother and can improve the stability of training.</p>