High-performance multi-scale target detection method based on deep learning

Abstract

The invention discloses a high-performance multi-scale target detection method based on deep learning, and the method comprises a training process and a detection process, and the training process comprises the following steps: 1.1, inputting a picture, and generating an image block; 1.2, screening positive image blocks; 1.3, screening negative image blocks; 1.4, inputting image blocks, and training a model; the detection process is as follows: 2.1, predicting a focus pixel set; 2.2, generating a focus image block; 2.3, a RoI stage; 2.4, carrying out classification and regression; 2.5, carrying out focus synthesis. According to the method, a brand new candidate region selection method is provided for the training process, meanwhile, a shallow-to-deep method is adopted for the detection process, regions which cannot possibly contain targets are ignored, and compared with a conventional detection algorithm for processing the whole image pyramid, the calculation amount of the multi-scaledetection method is remarkably reduced, and the detection accuracy is improved. The detection rate is greatly improved, and the bottleneck that the conventional multi-scale detection algorithm cannotbe put into practical application is broken through.

Description

technical field [0001] The invention relates to the field of information technology, in particular to a high-performance multi-scale target detection method based on deep learning. Background technique [0002] From the recognition of vehicles to the recognition of wearing masks during the epidemic, various target detection algorithms are widely used in today's society, such as Fast RCNN (fast convolutional network method based on region recommendation), Faster RCNN, YOLOv3, YOLOv4. The two most important properties to measure object detection are accuracy (mAP) and speed (FPS). However, the existing target detection algorithms either focus on the results of mAP, such as the two-stage algorithm of the Faster RCNN series, which has high precision, shares the amount of calculation, but is slow, takes a long time to train, and has a high false alarm rate. Either it emphasizes the balance between accuracy and speed, such as the one-stage algorithm of the YOLOv4 series, which is...

AI Technical Summary

Problems solved by technology

The same situation occurs in the detection process, which eventually leads to too many candidate regions, and the quality of most candidate regions also has problems, which reduces the detection speed

Images