Target detection method and system based on improved DAB-DETR

• Main Authors: ZHANG CHUANG, KATSURAHIKO
• Format: Patent
• Language: Chinese; English
• Published: 12.04.2024
• Subjects: CALCULATING; COMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS; COMPUTING; COUNTING; PHYSICS

The invention discloses an improved DAB-DETR-based target detection method and system, which is improved based on a DAB-DETR structure, develops two training designs on the basis of not changing an original structure, and effectively solves the problems of redundancy prediction between queries and misalignment between classification confidence and positioning precision. Specifically, a query-aware region constraint is proposed, and unnecessary internal conflicts between queries are avoided by limiting the responsibility region of each query. Then, a classification loss of position modulation is introduced, and the correlation between the classification confidence and the positioning precision is enhanced to encourage the model to generate the classification confidence to align with the prediction of the positioning precision. The provided two designs greatly improve the convergence speed and detection precision of the DAB-DETR, do not increase the parameter quantity and calculation cost of the model, can be e