A Weakly Supervised Object Localization Approach Using Convolutional Neural Networks to Correct Gradients

Abstract

The invention discloses a weakly supervised target positioning method using a convolutional neural network to rectify gradients, and relates to the technical field of computer vision. A convolutional neural network for classification function is trained on a given data set that only contains category labels. First, the network is forward-passed, and then the category of the target to be located is specified, and the convolutional neural network is corrected for the gradient. Reverse transmission, that is, the gradient is reversely transmitted layer by layer from the output layer to the input layer, and the corresponding correction operation is performed. The reverse transmission of the correction gradient of the convolutional neural network includes correction of the gradients transmitted by the fully connected layer and the convolutional layer in the network. The target in the generated heat map has a clear outline, and the obtained positioning accuracy is high. At the same time, it can distinguish different types of targets, and the localized area contains less irrelevant background. Robust to models with negative-valued features.

Description

technical field [0001] The invention relates to the technical field of computer vision, in particular to a weakly supervised target localization method using a convolutional neural network to correct gradients. Background technique [0002] In the field of computer vision, object localization using convolutional neural networks has achieved great success. However, a large category of existing target localization methods are based on supervised target localization methods. These methods require a large amount of labeled data to train convolutional neural networks. The training data needs to label the target category and target location information, especially the target location. Information requires a lot of human and material costs. Another type of method is based on weakly supervised target localization methods, for example, only using the category label information of the target to train a convolutional neural network for classification tasks, and then using the internal...

AI Technical Summary

Problems solved by technology

At present, the positioning heat map obtained by this kind of target positioning method based on weak supervision either contains a lot of noise, or cannot distinguish different targets, resulting in positioning accuracy far lower than that of the supervised target positioning method.

Images