Gray-scale image applicable neural network learning method and training method

Abstract

The invention proposes a gray-scale image applicable neural network learning method and a training method, comprising: preprocessing a gray-scale image as a second matrix; generating a binarized random coding matrix and then multiplying the second matrix as a fourth matrix; activating the function adjustment as a sixth matrix; establishing a binarized seventh matrix and an eighth matrix of floating point numbers; timing the seventh matrix by the fourth matrix as a ninth matrix; obtaining a tenth matrix representing the characters; subtracting the tenth matrix from the ninth matrix as an eleventh matrix; using the transpose matrix of the sixth matrix as a twelfth matrix; timing the sixth matrix by the twelfth matrix for a process parameter; dividing the twelfth matrix by the process parameter as a thirteenth matrix; timing the eleventh matrix by the thirteenth matrix as the fourteenth matrix; adding the fourteenth matrix to the eighth matrix to obtain a fifteenth matrix served as a new eighth matrix; and binarizing the fifteenth matrix as a new seventh matrix. According to the invention, the quantity of bytes in the matrix computation is reduced, the computation speed is increased and the hardware demand is lowered.

Description

technical field [0001] The invention relates to the field of neural networks, in particular to a neural network learning method and training method suitable for grayscale pictures. Background technique [0002] With the continuous evolution of computer and information technology, machine learning and pattern recognition have become one of the hottest fields in recent years. Some image recognition tasks that used to be performed by humans are gradually being replaced by machines, such as license plate recognition, face recognition, and fingerprint recognition. Although there are relatively mature solutions in these fields, the application fields of the solutions are very limited, and the expected recognition effect can only be achieved under certain conditions; in addition, traditional image recognition technology can only extract The partial information of the picture, but cannot identify and classify all the information in the test picture. It has a wide range of applicati...

AI Technical Summary

Problems solved by technology

Although there are relatively mature solutions in these fields, the application fields of the solutions are very limited, and the expected recognition effect can only be achieved under certain conditions; in addition, traditional image recognition technology can only extract Partial information of the picture, but cannot identify and classify all the information in the test picture. It has a wide range of applications and high recognition accuracy.

In the existing neural network learning method, due to the calculation involving multi-digit floating point numbers, the general calculation amount is too large, which slows down the calculation speed, so the requirements for computing equipment are very high

Images