Image super-resolution convolutional neural network computation acceleration method

Abstract

The present invention discloses an image super-resolution convolutional neural network computation acceleration method. The method comprises the following steps that: (1) a trained convolution kernelset is acquired; (2) the trained convolution kernel set is transformed into a convolution kernel set which is in a matrix form and is easy to be subjected to convolution computation and processing; (3) a convolution kernel set in an intermediate convolutional layer matrix form is obtained through analysis, so as to be adopted as an original convolution kernel set; (4) a low-rank learning model isconstructed based on the original convolution kernel set; (5) a base convolution kernel set is solved through the low-rank learning model; (6) a reconstruction coefficient is solved through using a least-squares model; and (7) convolution calculation equivalent substitution is performed on the convolution calculation of the original convolution kernel set through using the base convolution kernelset and the corresponding reconstruction coefficient, so that convolution calculation acceleration can be realized. According to the method of the invention, the convolution kernel set is reconstructed, and convolution computation acceleration is realized with an accuracy rate not decreased. The method only involves a convolution calculation process without changing an original accuracy rate, andcan be further combined with other acceleration methods to realize acceleration.

Description

technical field [0001] The invention relates to the field of performance optimization of convolutional neural networks (CNNs) accelerated calculation in artificial intelligence technology, in particular to an image super-resolution convolutional neural network accelerated calculation method. Background technique [0002] At present, convolutional neural network has become a popular technical method in the fields of computer vision, natural language processing and speech recognition, and has achieved huge technological breakthroughs. However, in the process of image super-resolution operation, due to the complexity of the calculation of the convolution operation itself and the deeper and deeper network layers, the calculation of the convolution layer takes a lot of time. rate presents a formidable challenge. Although the convolutional computing framework designed based on the GPU cluster can largely meet the training requirements of the deep convolutional neural network, the...

AI Technical Summary

Problems solved by technology

However, in the process of image super-resolution operation, due to the complexity of the calculation of the convolution operation itself and the deeper and deeper network layers, the calculation of the convolution layer takes a lot of time. rate poses a huge challenge

Although the convolutional computing framework designed based on the GPU cluster can largely meet the training requirements of the deep convolutional neural network, the application phase is usually based on mobile terminals or embeddable devices with limited local computing capabilities. It is difficult to meet the computing needs of large-scale networks, and at the same time, the shortest possible computing time is required in the application stage. For example, the super-resolution of pictures on the mobile phone has very high real-time requirements.

Therefore, the problem of accelerating convolutional computing has been restricting the application of convolutional neural networks.

Its shortcoming is that this method directly obtains low-rank sub-tensors and factor matrices through the tensor decomposition form of the original convolutional layer, which is a process of tensor approximation and cannot completely reconstruct the original convolution. layer, it is bound to cause a reduction in accuracy, especially if the effect reduction is more obvious for image super-resolution applications

Its shortcoming is that in this method, in the low-rank decomposition operation, the low-rank decomposition of the k*k convolution kernel in the original convolution layer is decomposed into two one-dimensional convolution kernels of k*1 and 1*k to perform convolution sequentially. In this way, since the reconstruction error of the original convolution kernel is not considered, the result after convolution is quite different from the original, and it needs to be retrained later to remove redundant connections in the network, and perform k-means clustering on the weights of the remaining connections As well as steps such as encoding the number of layers, although the acceleration is more obvious, the steps are too cumbersome, and the above method is a convolution acceleration for handwritten Chinese recognition. If the image super-resolution rate is applied, the accuracy rate will drop significantly.

The currently disclosed patents are not able to accelerate the convolution calculation for image super-resolution without reducing the PSNR index.

Images