Weak supervision target detection method based on positive and negative sample equalization

Abstract

The invention discloses a weak supervision target detection method based on positive and negative sample equalization. The method comprises steps of collecting scene images needing target detection, wherein the scene images correspond to labels, and a training set is mainly formed by the scene images and the corresponding labels; inputting the training set into a target candidate box screening module, and outputting an initial target candidate box corresponding to the scene image by the target candidate box screening module; establishing a weak supervision target detection network, inputting the training set and the corresponding initial target candidate box into the weak supervision target detection network at the same time for training, and obtaining the trained weak supervision target detection network in the training process; and inputting a to-be-detected scene image into the screening target candidate box module and the trained weak supervision target detection network, and classifying and positioning a target of the to-be-detected scene image. According to the method, only the image-level category label data set corresponding to the image is needed, and the good target detection capability is achieved under the condition that no target frame-level label exists.

Description

technical field [0001] The invention belongs to the technical field of computer vision, relates to a detection method of an image target, in particular to a weakly supervised target detection method based on positive and negative sample balance. Background technique [0002] The purpose of the object detection task is to locate and classify the objects existing in the image to be recognized. Since object detection is widely used in autonomous driving, video surveillance, industrial inspection, etc., this task has attracted extensive attention and research from industry and academia in recent years. [0003] In the target detection task, the fully supervised method requires the label at the target frame level to provide location and category information, and it is time-consuming and laborious to label a large number of images at the target frame level, and there is no dedicated large-scale labeling in many real-world application scenarios. Data sets, labeled data are very sc...

AI Technical Summary

Problems solved by technology

[0006] In order to solve the problems existing in weakly supervised target detection, the present invention provides a weakly supervised target detection method based on the balance of positive and negative samples, using the OICR network commonly used in weakly supervised target detection as the basic network model of the present invention, here Basically, it focuses on the imbalance of the number of positive and negative samples of target candidate frames in weakly supervised target detection and the existence of multiple objects of the same category, uses the information provided by weakly supervised semantic segmentation to screen negative samples, and uses the information in the middle of the training process to mine more Positive samples to improve the detection ability of the weakly supervised target detection model

Images