A graph-based top-down visual attention model for lockwire detection via multiscale top-hat transformation

As a locking device, lockwire is widely used in industrial fields that produces high levels of vibration. Lockwire detection is vital for ensuring mechanical stability. However, the existing methods are inapplicable to detect all types of small-size lockwires in complex backgrounds. In this paper, we construct a graph-based top-down visual attention model via the multiscale top-hat transformation to pick out lockwires from various complex backgrounds. Since lockwires typically exhibit separated bright spot-like structures at different scales in images, we firstly construct multiscale anisotropic Gaussian structuring elements to obtain the top-hat feature map. Based on a novel connectivity function, an undirected graph is then constructed. Afterwards, we propose an improved shortest path algorithm to remove prominent complex background components and extract lockwire candidates by minimizing the redefined cost function. Taking full advantages of imaging characteristics of lockwires, we design three saliency metrics to strengthen the saliency of lockwires while weakening backgrounds. Finally, a global top-down saliency map is produced to detect lockwires from complex backgrounds by combining three saliency maps. The experimental results show that our proposed method achieves superior performance compared to state-of-the-art methods.