| CPC G06V 30/19127 (2022.01) [G06T 7/12 (2017.01); G06T 7/66 (2017.01); G06T 7/73 (2017.01); G06V 10/82 (2022.01); G06V 30/19173 (2022.01); G06T 2207/20021 (2013.01); G06T 2207/20076 (2013.01); G06T 2207/20084 (2013.01)] | 9 Claims |
|
1. An English char image recognition method applied to an electronic device with computing capabilities, the electronic device comprising a processing unit, the English char image recognition method at least comprising steps of:
loading an English character image, generating a rectangular coordinate frame for the English character image by the processing unit;
finding a gravity center of the English character image in the rectangular coordinate frame, radiating 12 straight lines in sequence from the gravity center toward a direction of the rectangular coordinate frame at an interval angle of 30°, defining a simulated distance on each of the straight lines and calculating 3 vector feature positions from the simulated distance in order to obtain 36 vector features in the rectangular coordinate frame and read a gray scale of the vector feature positions;
dividing into four equal parts by 5 dividing points respectively on upper and lower edges of the rectangular coordinate frame and defining the dividing points as positions of upper and lower edge features to obtain 5 edge features respectively, dividing into five equal parts by 6 dividing points respectively on left and right edges of the rectangular coordinate frame and defining the dividing points as positions of left and right edge features to obtain 4 edge features in order to obtain 18 edge features on the rectangular coordinate frame and read a gray scale of positions of the edge features;
merging 36 vector features and 18 edge features and arranging in a 1×54 array, adding a blank feature respectively at beginning and end of the 1×54 array, arranging and forming a feature map in a 1×56 array;
performing 6 times of one-dimensional convolutional operation to generate 6 first array feature maps with an array of 1×54;
performing a maximum pooling operation of convolutional neural network on each of the first array feature maps to generate 6 second array feature maps with an array of 1×27;
performing a one-dimensional convolution operation of convolutional neural network on each of the second array feature maps to generate 16 third array feature maps with an array of 1×25;
performing a full connection of convolutional neural network on the third array feature maps to form a fourth full connection layer/fourth array feature maps with an array of 1×300;
performing a full connection of convolutional neural network on the fourth full connection layer/fourth array feature maps to form a fifth full connection layer/fifth array feature maps with an array of 1×150;
performing a full connection of convolutional neural network on the fifth full connection layer/fifth array feature maps to form a sixth full connection layer/sixth array feature maps with an array of 1×52; and
the processing unit performing probability recognition according to the sixth full connection layer/sixth array feature maps with an array of 1×52 and outputting 52 probabilities, and among the outputted 52 probabilities, outputting a class with a probability value floating point number being closer to 1 as a character recognition result.
|