Target detection method based on Transform and cascade characteristics

• Main Authors: TIAN CHENGJUN, LIU HAOBO, LIU ZHE, WANG YUYU, ZHANG JINTONG, YAN YU
• Format: Patent
• Language: Chinese; English
• Published: 25.07.2023
• Subjects: CALCULATING; COMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS; COMPUTING; COUNTING; PHYSICS

A target detection method based on Transform and cascade features relates to the technical field of computer vision technology, solves the problem that the existing detection precision needs to be improved, and comprises the following steps: dividing an original image into a plurality of image slices through a Patch segmentation module; inputting the image slices into a four-stage feature extraction backbone network formed by a shift window-based TransformerBlock module to obtain four feature maps with different scales from shallow to deep; inputting the feature maps into a cascade feature fusion module to perform feature map fusion among different scales to obtain five fused feature maps of different scales; extracting candidate frames according to the fused feature map to obtain candidate frames in which targets possibly exist; and inputting the fused feature map and the candidate box in which the target possibly exists into a target detection head based on ROIAlign to obtain a target detection result. Acco