An object detection system and method for active sample mining with dynamic selection strategy

Abstract

The invention discloses an object detection system and method for active sample mining with a dynamic selection strategy. The system includes: a sample acquisition unit for acquiring a small number of labeled samples and a large number of unlabeled samples; a model establishment and initialization unit for establishing Deep learning object detection model, and use a small number of labeled samples to initialize the deep learning object detection model; self-learning unit, used to introduce self-learning courses to guide the self-learning process to mine high-confidence samples from massive unlabeled samples to automatically perform pseudo-labeling ; The active learning unit is used to introduce the active learning course to guide the active learning process and mine the low-confidence samples in the massive unlabeled samples for manual labeling; the dual dual course constraint unit is used to introduce the dual dual course to guide the selection strategy in the self-learning process and the active learning process. Seamless switching is achieved during the learning process; the model training unit is used to train the model with selected pseudo-labeled samples and manually labeled samples to improve model performance.

Description

technical field [0001] The invention relates to the technical field of object detection and deep learning, in particular to an object detection system and method for active sample mining with dynamic selection strategy. Background technique [0002] In recent years, with the rapid development of the Internet and the improvement of computing power, deep learning technology that benefits from large-scale training data has made breakthrough progress. Object detection is a classic task in the field of computer vision, and the detection accuracy has been greatly improved. The object detection network based on candidate regions extracts a large number of candidate object regions from an input image, and performs category labeling and position regression on these regions, which improves the recognition accuracy and recognition speed of object detection. However, the detection effect of the neural network is extremely dependent on the labeled sample data of the training network. Th...

AI Technical Summary

Problems solved by technology

Therefore, manual labeling of images is very time-consuming and labor-intensive, and the development of methods for automatically labeling unlabeled data is a key step to reduce the burden of manual labeling

Images