When the watershed algorithm is applied to the infrared image segmentation of power equipment, the presence of image noise and gray-level variations caused by complex surface textures can lead to over-segmentation. To address this issue, an improved marked watershed algorithm combined with a K-means algorithm is proposed. First, the infrared image is preprocessed to suppress noise, and then combined with the gray-level information in the image. The equipment is extracted using the K-means clustering algorithm, and the resulting image is morphologically marked using an extended extremum transform based on the Otsu algorithm. Finally, the gradient image generated from the K-means clustering result is modified using the marked results to obtain the input for the watershed algorithm and complete the final segmentation. Experimental results show that the proposed method effectively reduces the sensitivity of the watershed algorithm to noise and gray-level variations, thereby overcoming the over-segmentation problem. Compared with the Otsu algorithm, region growing algorithm, and other classical methods, this approach segments only the external contours of the equipment while ignoring surface texture details.