Single-stage small sample target detection method for decoupling measurement

Abstract

The invention belongs to the field of computer vision, particularly relates to a single-stage small sample target detection method for decoupling measurement, and aims to solve the problems that an existing small sample detection target detection method is insufficient in detection precision, classification and regression interfere with each other in a non-decoupling mode, and detection network training is prone to over-fitting in the case of small samples. The method comprises the following steps: acquiring a to-be-detected image as an input image; obtaining a category and a regression box corresponding to each to-be-detected target in the input image through a pre-constructed small sample target detection network DMNet; wherein the DMNet comprises a multi-scale feature extraction network, a decoupling expression conversion module, an image-level metric learning module and a regression frame prediction module. According to the method, the overfitting problem during detection network training is avoided, the mutual interference of classification branches and regression branches is eliminated, and the precision of small sample target detection is improved.

Description

technical field [0001] The invention belongs to the field of computer vision, and in particular relates to a single-stage small sample target detection method of decoupling metrics. Background technique [0002] In the rapidly emerging field of deep learning, thanks to the training and learning of big data, deep neural networks have achieved great visual success. However, in some real-world scenarios, relevant visual data are scarce, such as underwater data and medical data, etc. In computer vision tasks, Convolutional Neural Networks (CNNs) have been widely used. The two-stage detector represented by Faster RCNN proposed by Ren et al. and the Single Shot MultiBox Detector (SSD) proposed by Liu et al. ) The single-stage detectors represented by ) have achieved good results in target detection. However, there is a problem in both types of detectors, that is, when the training data is very small, the training is prone to overfitting, which leads to a sharp drop in the perfor...

AI Technical Summary

Problems solved by technology

[0005] In order to solve the above problems in the prior art, that is, in order to solve the insufficient detection accuracy of the existing small sample detection target detection method, the mutual interference of classification and regression in the non-decoupled form, and the detection network training is prone to overfitting in the case of small samples In the first aspect of the present invention, a single-stage small-sample target detection method for decoupling metrics is proposed, which includes the following steps:

Images