Development of a deep learning image classification network based on vehicular collision using instance mask guided attention

For successful navigation of autonomous vehicles, on-road vehicular collision understanding is significant. In the present study, an attempt has been made to develop an Instance Mask-Guided Attention (IMGA) network for image classification based on vehicular collision, by supervision of the learnt attention map with the generated masks of colliding vehicles. This contributes to the development of a differential learning method that focuses just on the image's significant parts, thereby boosting its capability. The main contributions of this study are the creation of a vehicular collision image classification dataset and the development of a modified Visual Geometry Group-19 network with IMGA that achieved 97.94% accuracy in the supplied dataset, outperforming various current deep learning-based image classification techniques. Extensive experimentation demonstrates that a simpler feature extraction backbone is adequate to achieve higher image classification accuracy because of the effect of instance mask guidance, if not a deeper feature extraction network with increased computational complexity would be required.