A method, device and medium for image recognition based on pulsation array

Abstract

The embodiment of the present invention discloses an image recognition method, device and medium based on a pulsation array, which converts the acquired image feature information into a one-dimensional feature vector; converts the acquired weight matrix into a one-dimensional weight vector, and uses the trained Each node in the 3D systolic array model is assigned a corresponding weight group. Using the three-dimensional systolic array model, the one-dimensional feature vector is multiplied and accumulated by the weight value in parallel to obtain the corresponding feature value of each node. The feature values ​​with different values ​​can reflect the category of items contained in the image, and the category of items contained in the image can be determined according to the feature value corresponding to each node and the corresponding relationship between the feature value and the item category established in advance. After the image feature information and weight value to be calculated are transformed into one dimension, the structure of accelerated calculation using the pulsating array model is used to calculate the one-dimensional feature vector, which fully expands the parallelism of vector calculation and effectively improves the image classification processing efficiency.

Description

technical field [0001] The present invention relates to the technical field of image processing, and in particular, to an image recognition method, device and computer-readable storage medium based on a systolic array. Background technique [0002] At present, the research on deep learning mainly focuses on Convolutional Neural Network (CNN) as the research object. Due to different processing scenarios, the performance requirements of CNN are also different, so a variety of network structures have been developed. But the basic composition of CNN is fixed, which are input layer, convolution layer, activation layer, pooling layer and fully connected layer. [0003] Fully connected layers (FC) play the role of "classifier" in the entire convolutional neural network. Operations such as convolutional layers, pooling layers, and activation layers map the original data to the hidden layer feature space, while the fully connected layer plays the role of mapping the learned “distri...

AI Technical Summary

Problems solved by technology

[0004] At present, there are mainly two methods for calculating the fully connected layer. One is to calculate the fully connected layer according to the convolution with a convolution kernel size of 1*1, but it is necessary to replace the data of the convolution kernel, that is, the weight value very quickly. The bandwidth pressure of data transmission is very high

The second is to use the method of all or part of the parallel calculation of the elements corresponding to the feature value and the weight value. Because of the parallel calculation, an additional intermediate result processing link is required, so it cannot be continuously calculated, which reduces the use efficiency of hardware resources and affects the classification of images. efficiency

Images