Research on the multispectral image fusion and shelf status recognition in UAV inspection

With the advancement of intelligent inspection technologies, unmanned aerial vehicles (UAVs) have seen increasing applications in industries such as power, agriculture, and logistics. However, traditional manual inspections in warehousing and logistics suffer from low efficiency and high rates of false or missed detections, necessitating improvements in the automation and accuracy of shelf status recognition. This study aims to enhance the multidimensional feature representation of images by fusing multispectral image information—including visible light, infrared, and thermal imaging—to enable comprehensive detection of shelf item presence, category identification, and temperature anomalies. A UAV-based inspection and data acquisition system was built in a real warehouse environment to collect a multispectral image dataset under varying times, angles, and lighting conditions. Through image registration and multi-scale feature fusion techniques, spatial alignment and information enhancement across image modalities were achieved. Experimental results show that the proposed MCANet model achieved an average recognition accuracy of 94.6% on the self-constructed dataset, an improvement of 8.3 percentage points over traditional visible-light-only models. For the task of missing item detection, the fusion model achieved a recall rate of 92.1%, significantly outperforming the baseline YOLOv5 model's 83.7%. The main innovations of this study include the collection and public release of a high-quality multispectral image dataset for UAV shelf inspections, providing valuable data support for related research. Additionally, a multispectral recognition model incorporating attention mechanisms is proposed, enhancing the extraction of key targets in complex scenes.