Winograd algorithm-based rapid image processing method

Abstract

The invention discloses a winograd algorithm-based rapid image processing method, which comprises the following steps of: step 1, selecting a data set, training a self-defined neural network model byutilizing a Caffe framework, and extracting a convolution kernel weight and a bias value of the trained model; step 2, extracting input picture pixel points, and storing the input picture pixel pointsin a four-dimensional array, the four dimensions being the number of input pictures, the number of channels, and the length and width of the pictures respectively; step 3, constructing a convolutionoperator based on a winograd algorithm, judging whether the convolution kernel size is 3 * 3 and whether the channel number is greater than 10, and if so, performing convolution operation by using thewinograd operator; and step 4, outputting a result obtained after the convolution operation, judging whether the layer is the last convolution layer or not, if so, sending the output picture into thefull connection layer after nonlinear transformation of the RELU layer, and otherwise, repeating the step 3. The image processing method can improve the computing energy efficiency when the processorruns the neural network.

Description

technical field [0001] The invention belongs to the field of embedded image recognition, in particular to a fast image processing method based on a hybrid operator combining a winograd operator and a traditional convolution operator. Background technique [0002] With the rapid development of software and hardware technology and the advent of the era of big data, deep learning technology has achieved breakthroughs in more and more fields, and the typical application field is image recognition. Image recognition is to compare the stored existing information (the stored information in the memory unit) with the current input information (the current information received by the senses), use the computer to process the image, analyze and understand the context, and identify various objects Technology for objects or objects. Compared with traditional image processing systems such as image processing cards, embedded image processing systems have the advantages of small size, low c...

AI Technical Summary

Problems solved by technology

However, in the current popular neural network framework, the massive parameters in the traditional convolution operator bring a large burden to the processor. For example, in VGGnet, the convolution parameter value reaches more than 100M, which greatly increases the Computational overhead

[0005] In order to overcome some deficiencies in neural networks, such as too many parameters required by traditional convolution operators, and excessive computing resource overhead, this case arises from this

Images