Exploration of convolutional neural networks in automated animation scene production

In animation production, scene colorization is time-consuming and labor-intensive, necessitating intelligent solutions. To address automatic animation scene colorization, this paper proposes a hybrid Cross-CNN model based on convolutional neural networks, which extracts local features through ResNet50 network and combines Transformer architecture to obtain global information, achieving automatic colorization of animation sketches. A cross-attention mechanism is designed to enhance feature matching capability and improve color consistency between different frames. Experiments conducted on a dataset containing 20,000 pairs of samples show that the model achieves an SSIM index of 0.932, representing a 1.4 percentage point improvement over existing methods. Through ablation experiments, the effectiveness of the cross-attention mechanism is validated, providing a new technical approach for automated animation scene production.