MULTI-KERNEL IMAGE PROCESSING CONVOLUTIONAL NEURAL NETWORK-ORIENTED DATA READING METHOD

摘要

A multi-kernel image processing convolutional neural network-oriented data reading method comprises: A) receiving a convolution computation parameter comprising an image size, a channel number, a convolution kernel size K x K, and/or a step size; B) determining a segmentation operation count for each row of data of an image and a length of remaining data from the last operation, such that each image block obtained from segmentation yields data required for M convolution operations; C) performing image segmentation according to a determined width of a segmented image block, reading, for a first channel of an image block, first K rows of data of the segmented image block, and storing the same in a data buffer unit; D) reading a first row of image data stored in the data buffer unit, and expanding the first row of data of each first channel in the previous M convolution operations to generate M*K pieces of data in total; E) reading and expanding the second row to the Kth row of the image data stored in the data buffer unit; F) repeating step C) and step E) for a second channel to a last channel; G) a data reading unit returning to the first channel and reading a next row of data, overriding data in a row having a frontmost serial number in the first channel in a buffer, and executing step D) and step E) on the updated K rows of data in the data buffer unit; H) repeating step G) for the second channel to the last channel; and I) repeating step G) and step H) until the last row of the image block is completed.