A target recognition method based on smooth multi-instance learning

Abstract

The invention relates to a target recognition method based on smooth multi-instance learning, which includes a sample training step and a target recognition step. The sample training step is: extracting example features of training pictures; performing smooth processing according to the example features to obtain an example-related continuous probability model ;Training according to the example-related continuous probability model to obtain the training weight w; the target recognition step is: extract the example features of the picture to be detected; according to the example features and training weights of the picture to be detected, calculate and determine whether there is a need to identify The goal. Compared with the prior art, the present invention has the advantages of high recognition efficiency, high recognition accuracy and good recognition effect.

Description

technical field [0001] The invention relates to the field of image recognition, in particular to a target recognition method based on smooth multi-instance learning. Background technique [0002] Multi-instance learning was proposed by Dietterich et al. in 1997, and it was mainly used in the prediction of drug activity at that time. Since then, due to the wide application of multi-instance learning in machine learning and machine vision, a large number of multi-instance learning methods have been proposed. In 1998, DD (Diverse Density) was proposed. This method deals with multi-instance learning problems by finding blocks in many different positive packets and a small number of negative packets. Later, this method was developed into DD-SVM. In 2002 and 2005, S.Andrews and C.Zhang proposed the miSVM and MILBoost methods respectively, and they classified examples by training SVM and boosting classifiers. Recently, in 2015, J.Wu et al. established a deep learning framework an...

AI Technical Summary

Problems solved by technology

[0004] (1) Supervised image target recognition relies heavily on human participation. It is necessary to manually mark whether the image contains the target object, and also mark the position of the object in the picture containing the object.

In the case of a large sample size, this is very time-consuming and laborious, and there are certain limitations

[0005] (2) Not all multi-instance learning methods can identify the labels of examples in a bag, and some multi-instance learning methods only classify the bag-level samples

[0006] (3) In the traditional multi-instance learning method, the detector used to locate the target object in the positive sample is trained on the same sample, which makes it biased when relocating in the same window

[0007] The above three methods have failed to achieve good recognition results due to their own defects.

Images