Deep learning image target mapping and positioning method based on weak supervision information

Abstract

The invention provides a deep learning image target mapping and positioning method based on weak supervision information. The method comprises the following steps that: using image data with a category tag to independently train two deep convolutional neural network frames to obtain a classification model M1 and a classification model M2, and obtaining the parameters of a global learnable poolinglayer with parameters; using the new classification model M2 to carry out characteristic extraction on a test image to obtain a characteristic pattern, and according to the characteristic pattern, obtaining a preliminary positioning box through characteristic category mapping and a threshold value method; using a selective searching method to carry out candidate area extraction on the test image,and using the classification model M1 to screen a candidate box set of which the category belongs to a target object; and carrying out non-maximum inhibition processing on the preliminary positioningbox and the candidate box to obtain the final target positioning box of the test image. By use of the method, the global learnable pooling layer with parameters can be imported, learning can be carried out to obtain a better characteristic expression about a target category j, and a selective characteristic category mapping way is used for effectively obtaining the position information of the target object in the image.

Description

technical field [0001] The invention relates to the technical field of image processing, in particular to a deep learning image target mapping and positioning method based on weakly supervised information. Background technique [0002] With the development of deep learning technology represented by deep convolutional neural network, great breakthroughs have been made in the fields of image classification and image object recognition, and many influential academic researches and related industrial applications have been triggered. In the 2015 Global Large-Scale Visual Recognition Competition (ILSVRC), the deep residual model proposed by Microsoft Research Asia won the championship with a recognition error rate of 3.57%, and surpassed human recognition accuracy for the first time. [0003] The Regional Convolutional Neural Network (RCNN) proposed in 2014 was the first to use a deep convolutional network for image target detection tasks, and its performance was significantly im...

AI Technical Summary

Problems solved by technology

[0006] The disadvantage of the above-mentioned image mapping method in the prior art is: this method of pooling the feature map using the method of calculating the global average value or the global maximum value will cause the feature points after pooling to have insufficient representation ability for the original features.

Images