Results and Discussions In the contour extraction method for reference holes, this paper proposes saliency detection as an image preprocessing method to avoid the poor image quality caused by objective factors. Such factors pose great difficulty to the accurate extraction of the contour of the reference hole by the traditional threshold segmentation and Canny edge detection algorithms and improving the robustness of reference hole identification and detection. In terms of the accuracy of reference hole positioning, this paper proposes the method of performing double-threshold segmentation and principal component analysis to coarsely position the contour of the reference hole. Then, the Bazen method is used to extract the sub-pixel contour and further accurately position the contour of the reference hole. Finally, the RANSAC method is adopted to fit the contour data and thereby improve the accuracy of ellipse fitting. According to experimental verification, the positioning error of the noisy image is 0.202 pixel, and that of the real reference hole is 0.027 mm. Clearly, the detection performance is excellent, and the proposed method is still robust in reference hole detection under the conditions of reflection, cutting tool traces, uneven light, threaded holes with chamfers, embedded contours, occlusion, and so on.

The complex environment and poor image acquisition quality of aerospace assembly sites pose great difficulty to the detection and identification of assembly reference holes due to reflection, cutting tool traces, uneven light, threaded holes with chamfers, embedded contours, occlusion, and other conditions. To address the above problems, this paper proposes a technique of reference hole detection based on saliency detection. In the process of aerospace assembly, the traditional assembly methods relying on manual operations or special fixtures are transformed into digital and flexible assembly to improve assembly efficiency and ensure assembly accuracy. The references on components need to be measured before assembly. Although contact measurement offers high accuracy, it is slow and inefficient. Therefore, non-contact measurement is often chosen, and assembly components are measured by machine vision systems, namely that the positioning references on the components are measured by vision. The above methods are based on the acquisition of the edge points of the reference holes. Nevertheless, Canny edge detection, threshold segmentation, and other image preprocessing methods can hardly accurately extract the actual edge points of reference holes in actual working conditions due to reflection, cutting tool traces, uneven light, threaded holes with chamfers, embedded contours, occlusion, and other conditions. For the above reason, this paper assumes that the image can be transformed into a saliency map before detection and accurate contour positioning can be achieved by principal component analysis and sub-pixel extraction.

The traditional process of reference hole detection mainly consists of image preprocessing, contour extraction, and contour recognition. Among them, image preprocessing is to reduce noise and highlight the reference hole in the image. Contour extraction is to obtain all the contours in the preprocessed image, although a lot of them are not the contour of the reference hole. Contour recognition is to eliminate all such contours from all the contours to be detected to retain the contour of the reference hole and ultimately achieve accurate detection of the reference hole. Images taken at the actual assembly site indicate that the traditional detection methods can hardly accurately extract the contour of the reference hole, resulting in the false-positive and false-negative detection of the reference hole. To overcome the above problem, this paper preprocesses the image by saliency detection to transform the image into a saliency map, highlights the saliency of the reference hole in the image, and extracts the region of interest regarding the reference hole from the saliency map with the Hough circle detection algorithm. Subsequently, double-threshold contour segmentation is performed, and principal component analysis is conducted to extract the pixel-level contour of the reference hole. Then, the Bazen method is employed to extract the sub-pixel contour of the reference hole. Finally, the reference hole is positioned with high precision according to the principle of random sample consensus (RANSAC).

The proposed method of reference hole detection based on saliency detection can be applied to detect reference holes under reflection, cutting tool traces, uneven light, threaded holes with chamfers, embedded contours, occlusion, and other conditions, in which it still ensures the robustness of detection. Experimental verification shows that the positioning error of the noisy image is 0.202 pixel, and that of the real reference hole is 0.027 mm. The method has a bright application prospect in industrial assembly reference, processing, and the positioning hole of a task. Compared with existing methods, this method can detect reference holes that are difficult to detect for most of the traditional methods. On the basis of a known size of a reference hole to be detected, the minimum and maximum radius thresholds can be set for the Hough circle detection algorithm to guarantee the robustness of the proposed detection method.