Underground personnel recognition based on low-light enhancement of infrared and visible light image fusion

In underground environments, there is low light, and personnel features are not clearly visible. Existing infrared and visible light image fusion methods based on deep learning use only infrared information to fill in the scene defects caused by the light degradation of visible images under weak lighting conditions. As a result, rich scene information from visible images is lost in the fused image in dark environments. Moreover, treating image enhancement and image fusion as separate tasks leads to poor fusion results. To address the above issues, a model for underground personnel recognition based on low-light enhancement of infrared and visible light image fusion was proposed. First, the visible and infrared sensor images underwent preprocessing steps such as grayscaling and geometric correction. Then, the processed images were passed into a low-light enhancement network, which removed the illumination component from the degraded visible light images at the feature level. Finally, texture-contrast enhancement networks performed feature-level fusion, enhancing overall visual perception in terms of texture and contrast. Experimental results showed that the proposed model improved underground personnel recognition results compared to the visible light modality, with an average accuracy increase of 8.2%, recall rate increase of 12.5%, and [email protected] increase of 8.3%. Compared to the infrared modality, accuracy increased by an average of 2.1%, recall rate increased by 5.1%, and [email protected] increased by 4.1%. Meanwhile, the detection speed reached 31.2 frames/s, solving problems such as misdetection and missed detection caused by unclear personnel features in low-light underground scenarios.