In complex environments such as power grids and bridges, inspection robots can effectively replace manual inspections and equipment maintenance. To achieve autonomous obstacle avoidance for inspection robots in complex environments, this study proposes an image-matching recognition technique based on the YOLOv5 deep-learning network model, which utilizes both IR and visible-light images. This enables the robot to identify and classify various obstacles, including living and nonliving obstacles. During the inspection operations, the robot is equipped with an IR camera and a regular camera to monitor its environment in real time. With the YOLOv5 network model trained on a large dataset, the robot can quickly and accurately identify and categorize obstacles along its path. The robot not only identifies the nature of obstacles but also performs appropriate proactive actions to address different situations. The average recognition accuracy is approximately 99.2%. Experimental results demonstrate the effectiveness of the comprehensive obstacle avoidance method based on multiple-image information. The robot can detect, classify, and navigate obstacles under various scenarios, thereby enhancing their autonomy and adaptability. This technology has wide applications in areas such as automated inspections, safety monitoring, and rescue missions, providing strong support for the continuous development of robotic technologies.