Chinese-character-detection-and-recognition

With Yolov5 and a simple CNN, this project can detect and recogize black chinese characters on white paper.

With the detected boxes and predicted labels, a standard character with font simka.ttf can be easily generated, before the score is calculated.

source	standard
matched	total

The matched map shows the pixels that the tow maps above have in common, while total map is the result of adding them.

Than, the score is calculated. $$ score = \frac{number\ of\ matched\ pixels}{number\ of\ total\ pixels} $$

Usage

python run.py --save-txt --conf 0.25 --iou 0.1

If everything gose fine, you can get the output like this.

Namespace(agnostic_nms=False, augment=False, classes=None, conf_thres=0.25, device='', exist_ok=False, img_size=640, iou_thres=0.1, name='exp', project='runs/detect', save_conf=False, save_txt=True, source='single_word/imgs/', update=False, view_img=False, weights='single_word/weights/yolov5s.pt')
YOLOv5 181b015 torch 1.7.1+cu110 CUDA:0 (GeForce 940MX, 2048.0MB)

Fusing layers...
Model Summary: 224 layers, 7053910 parameters, 0 gradients, 16.3 GFLOPS
image 1/1 some_path_to\single_word\imgs\test.jpg: 384x640 Done. (0.066s)
Results saved to runs\detect\exp
1 labels saved to runs\detect\exp\labels
Done. (1.934s)

Similar to yolo's label ouput, the results in the above lable path is a txt file. With a score added, each row is class x_center y_center width height score format.
class is alwsay 0, since only one class named word is set.
Box coordinates are in normalized xywh format (from 0 - 1).

Result

original	processed

The standard character will be printed to corresponding area, with a confidence and a score for each detected boxes.