Detecting slender objects with uncertainty based on keypoint-displacement representation. (July 2021)

• Record Type: Journal Article
• Authors: Kong, Zelong; Zhang, Nian; Guan, Xinping; Le, Xinyi
• Is Part Of: Neural networks. Volume 139(2021)
• Page Start: 246
• Page End: 254
• Publication Date: 2021-07
• Subjects: Deep learning -- Object detection -- Uncertainty prediction -- Quality evaluation; Neural computers -- Periodicals; Neural networks (Computer science) -- Periodicals; Neural networks (Neurobiology) -- Periodicals; Nervous System -- Periodicals; Ordinateurs neuronaux -- Périodiques; Réseaux neuronaux (Informatique) -- Périodiques; Réseaux neuronaux (Neurobiologie) -- Périodiques; Neural computers; Neural networks (Computer science); Neural networks (Neurobiology); Periodicals; 006.32
• DOI: 10.1016/j.neunet.2021.03.024
• Languages: English
• ISSNs: 0893-6080

Abstract

Abstract: Slender objects are long and thin objects. Existing object detection networks are not specially designed for detecting slender objects. We propose a method to detect slender objects. We represent slender objects with a keypoint-displacement pattern instead of using axis-aligned bounding boxes, avoiding problems like orientation confusion and wrong elimination. In our network, three parallel branches predict keypoint heatmaps, displacement vector field, and displacement uncertainty heatmap respectively. We add the uncertainty branch to enable our network to give uncertainty together with detection results. The predicted uncertainty provides a continuous criterion to evaluate whether detection results are reliable. In addition, the uncertainty branch can lower the weight of ambiguous training samples, leading to more accurate detection results. We employ our proposed method in two typical practical applications. Edges of electrode sheets and pins of electronic chips are correctly detected as slender objects. Manufacturing quality is evaluated through analyzing the detection results, including keypoint number, displacement property, and uncertainty value.