High-precision and high-proportion classification model and classification method for realizing dynamic calculation

Abstract

The invention relates to the technical field of neural networks, and provides a high-precision and high-proportion classification model and classification method for realizing dynamic calculation, and the high-precision and high-proportion classification model comprises a trunk network model and at least one first branch network model; wherein the backbone network model comprises an input end, an output end and a plurality of intermediate layers, and a first branch network model is inserted between 1 / 4-1 / 3 layers of the plurality of intermediate layers; the first branch network model comprises a classification module and a confidence determination module. In the actual application process, the classification module is used for generating classification output at the insertion position, the confidence degree determination module is used for judging whether the classification output meets the precision requirement or not, if yes, the high-precision high-proportion classification model exits from the output end of the first branch network model in advance, and if not, the classification output generated by the classification module is given up, and it is returned to the middle layer of the backbone network model to continuously complete the classification operation.

Description

technical field [0001] The present application relates to the technical field of neural networks, in particular to a high-precision and high-scale classification model and classification method that realize dynamic calculation. Background technique [0002] With the development and maturity of neural network technology, more and more industrial fields use deep neural network technology to complete industrial tasks. For example, the classification neural network obtained through image data learning can classify industrial scenes or industrial products. Compared with Artificial, classification neural networks can not only complete the work efficiently, but also greatly improve the accuracy rate. [0003] However, in some special application scenarios, due to the large amount of parameters and calculation of the deep neural network, the power consumption and delay of deployment on the computing platform are relatively high, which greatly restricts the application scenarios. For...

AI Technical Summary

Problems solved by technology

[0003] However, in some special application scenarios, due to the large amount of parameters and calculation of the deep neural network, the power consumption and delay of deployment on the computing platform are relatively high, which greatly restricts the application scenarios. For this reason, many single compression A scheme or model complexity reduction scheme is proposed, such as quantization sparseness and dynamic calculation, etc.

[0004] However, in actual use, due to the complexity of industrial scenes, there will be simple samples and complex samples. For example, in the process of image recognition, multiple images of the same object will be different due to light, angle, image size and clarity. , the recognition difficulty of multiple pictures is not the same. The scheme of the prior art has the same recognition process when recognizing a class of pictures. Although the recognition time of simple pictures will be less than that of complex pictures, there will still be excessive Use on-chip resources to identify simple pictures, resulting in a waste of time and on-chip resources

Images