Font detection method and system using artificial intelligence-trained neural network
The present disclosure relates to a font detection method using a neural network. The font detection method using the neural network according to the present disclosure includes receiving a target text image including a text; resizing a horizontal or vertical size to a reference input size according to an aspect ratio of the input target text image; and inputting the resized target text image into a trained neural network and outputting a font of the text included in the text image, and the neural network may be trained with a unit image extracted as a unit region of the reference input size from a training image generated by synthesizing a background with the text. According to the present disclosure, fonts according to various usage examples may be effectively detected.
1. A font detection method using a neural network, the font detection method comprising:
receiving a target text image including a text;
resizing a horizontal or vertical size to a reference input size according to an aspect ratio of the input target text image; and
inputting the resized target text image into a trained neural network and outputting a font of the text included in the text image,
wherein the neural network is trained with a unit image extracted as a unit region of the reference input size from a training image generated by synthesizing a background with the text, and
wherein the neural network comprises a convolution layer that extracts features through a convolution operation on the input target text image, a pooling layer that extracts a value representing a feature for each channel with respect to an extracted feature map, and a fully connected layer that outputs a probability for each category through an output of the pooling layer.
2. The font detection method of claim 1 , wherein the receiving includes extracting a region including the text from an uploaded image by user and receiving the extracted region of the uploaded image as the text image.
3. The font detection method of claim 1 , wherein the receiving includes extracting a region of the text from an arbitrary web page including the text on a web and receiving the extracted region of the web page as the text image.
4. The font detection method of claim 1 , wherein the neural network is trained by resizing the horizontal or vertical size to the reference input size according to the aspect ratio of the training image, and extracting a square unit image of the reference input size from the resized training image.
5. The font detection method of claim 1 , wherein the outputting of the font includes
extracting feature information from the target text image input to the neural network by performing a convolution operation through a filter of a predetermined size;
extracting a representative feature for each channel by which the extracted feature information is defined; and
extracting corresponding font category information from the extracted representative feature.
6. The font detection method of claim 1 ,
wherein the receiving includes receiving a plurality of target text images in a batch unit, and
wherein the resizing includes padding a region of another image with a blank with respect to an image having a largest size resized according to the reference size among the plurality of images.
7. The font detection method of claim 1 , wherein the neural network dynamically sets a condition for giving an effect to the text or a condition for synthesizing a background according to a training result through a training data set including an image including a text created in an arbitrary font and the background, and the font used.
8. A non-transitory computer-readable recording medium having stored thereon a program for performing the font detection method using the neural network according to claim 1 .
9. A font detection device using a neural network, the font detection device comprising:
a text input unit receiving a target text image including text;
a resizing unit resizing a horizontal or vertical size to a reference input size according to an aspect ratio of the input target text image; and
a font output unit inputting the resized target text image into a trained neural network and outputting a font of the text included in the text image,
wherein the neural network is trained with a unit image extracted as a unit region of the reference input size from a training image generated by synthesizing a background with the text, and
wherein the neural network comprises a convolution layer that extracts features through a convolution operation on the input target text image, a pooling layer that extracts a value representing a feature for each channel with respect to an extracted feature map, and a fully connected layer that outputs a probability for each category through an output of the pooling layer.
10. The font detection device of claim 9 , wherein the neural network includes
a feature extraction unit extracting feature information from the target text image input to the neural network by performing a convolution operation through a filter of a predetermined size;
a feature representation unit extracting a representative feature for each channel by which the extracted feature information is defined; and
a font categorization unit extracting corresponding font category information from the extracted representative feature.
11. The font detection device of claim 10 , wherein the neural network dynamically sets a condition for giving an effect to the text or a condition for synthesizing a background according to a training result through a training data set including an image including a text created in an arbitrary font and the background, and the font used.
12. The font detection device of claim 9 ,
wherein the text input unit receives a plurality of target text images in a batch unit, and
wherein the resizing unit performs padding on a region of another image with a blank with respect to an image having a largest size resized according to the reference size among the plurality of images.