CAGNet: A Framework for Automotive Crash Test Scene Image Classification

In automotive crash tests, a large volume of images is generated, making manual classification inefficient and error-prone, particularly under multi-view conditions. To address these challenges, we propose CAGNet (ConvNeXt V2 + ALLMIX + GMNet) for automotive crash test image classification. CAGNet adopts ConvNeXt V2 as the backbone and incorporates an adaptive multi-scale feature fusion module (ALLMIX) together with a global–local feature cooperative modeling module (GM-Net), aiming to enhance representation capability across different collision stages, imaging viewpoints, and subtle structural variations. The proposed method is evaluated on a self-constructed automotive crash test image dataset, with additional generalization experiments conducted on the CIFAR-100, Tiny-ImageNet, and Flowers-102 datasets. Experimental results demonstrate that CAGNet consistently outperforms mainstream models, including ResNet-50, EfficientNet-B1, Swin-T, ConvNeXt-Tiny, and ConvNeXt V2-Tiny. On the self-constructed automotive crash test image dataset, CAGNet achieves improvements of 1.90, 0.52, and 1.84 percentage points in Top-1 accuracy, Top-5 accuracy, and F1-macro, respectively, compared with the strongest baseline, ConvNeXt V2-Tiny.