Enhancing Generalization in Deep Learning-Based Image Recognition under Complex Scenes

As a core technology in AI, deep learning (DL) can automatically learn features of images by using deep neural networks to achieve high image recognition accuracy. In order to address image recognition problems in complex scenes, a DL model with attention and multi-scale features is proposed. Incorporating the channel attention (CA) and spatial attention (SA), the model can target the critical areas in images and depress background noise as well as meaningless information. And it adopts feature pyramid network (FPN) form to achieve multiscale feature fusion, enhancing the recognition of object targets for various scales. The experiment analysis shows the effectiveness of our model compared with the typical VGG and ResNet models using CIFAR-10 and ImageNet data sets in accuracy, recall and F1 score. Particularly, in processing images in complex scenes, the model has better recognition and robustness, which verifies the effectiveness of the attention mechanism and the multiscale feature fusion strategy. Research findings not only improve the DL model's image identification performance in challenging scenes but also offer solid theoretical and technical justification for use in related domains.