Stochastic VI-TIR Image Fusion Method and Verification Using Lightweight Models

The recognition of only visible image is difficult to meet the rapidly developing application requirements of increasingly complex application scenarios in automatic driving and industrial automation. More and more researchers have turned their attention to the field of multi-modal image fusion. In this paper, a stochastic fusion method of multi-modal image using infrared image as mask is introduced. Considering the deployment requirements of edge computing platforms, the recognition rate of fused image and non-fused image is compared and verified on several lightweight models. The results show that compared with non-fused images, the fusion image provided by this method can improve the recognition rate by 5%∼7% and reduce the storage cost of these fused images greatly(uint8 reduced to uint3). In a meanwhile, the accuracy of classification models for the stochastic VI-TIR fusion image is roughly equivalent to that of fusion images acquired by TarDAL and LDFusion methods. The results demonstrates that the stochastic fusion method is conducive to multi-modal image fusion and local recognition on the edge computing platforms.