Detection method for image classification output results

Abstract

The invention relates to a detection method for image classification output results, which is to construct a first classification network with a multi-layer network by analyzing the network layer layer by layer; it includes a convolutional layer and a fully connected layer; a convolutional layer and a fully connected layer The original image is processed linearly and nonlinearly to obtain a feature map, and the feature map is converted to form introspection network training data; if the prediction result of the first classification network is wrong, the training label is 0; if the prediction result of the first classification network is correct, the training label is It is 1; input the training sample into the introspection network; input the output result of the first classification network into the trained introspection network after the processing and conversion in step S2, and judge whether the output result of the first classification network is correct through the output of the introspection network. The invention recognizes the output feature map of the first classification network through the introspection network, and then can judge whether the output result of the first classification network is correct.

Description

technical field [0001] The invention relates to the technical field of computers, in particular to a method for detecting output results of image classification. Background technique [0002] By learning translation-invariant features and parameter sharing, using convolutional neural networks (CNN, ConvolutionalNeural Networks) to perform image classification tasks greatly improves classification performance and reduces the number of learning parameters, making the network more generalized and less prone to overfitting . However, the performance of these networks is still highly dependent on the datasets they are trained on. A dataset that deviates from the norm can lead to incorrect predictions, mainly because the network has not seen similar images when it was trained. When these networks use a softmax to compress the output between [0,1] and force the scores to become probabilities that sum to 1, even though all categories have lower scores, due to the normalization exp...

AI Technical Summary

Problems solved by technology

[0004] In view of this, the purpose of the present invention is to overcome the deficiencies of the prior art and provide a detection method for the output result of image classification to solve the problem in the prior art that it is impossible to judge whether the output result of the classification image is accurate

Images