Simple Numerical KCAPTCHA solver using YOLOv8

1. Install requirements with below command

   pip3 install -r requirements.txt

2. Generate KCAPTCHA dataset

   1. Generate Image-label set with kcaptcha-generator
      • Recommand to generate >50K image-label set for training
   2. Split them into train, test, val set
   3. Reformat BBOX label to YOLO label format using bbox2yolo.py

      usage: bbox2yolo.py [-h] [--path PATH]
      
      options:
      -h, --help   show this help message and exit
      --path PATH  dataset root path

3. Configure your training config files as you want

4. Train YOLO with below command

   usage: train.py [-h] [--config CONFIG] [--log-path PATH] [--devices DEVICES] [--dataloader-workers WORKERS] [--img-size IMG_SIZE] [--model MODEL] [--epochs EPOCHS]
                   [--batch-size BATCH_SIZE] [--no-finetune] [--optimizer OPTIMIZER] [--val]
   
   options:
   -h, --help            show this help message and exit
   --config CONFIG       path to yolo config file
   --log-path PATH       Relative path to save model
   --devices DEVICES     Devices to use on training
   --dataloader-workers WORKERS
                           Dataloader workers count
   --img-size IMG_SIZE   max width of image
   --model MODEL         model to use
   --epochs EPOCHS       Epochs to train
   --batch-size BATCH_SIZE
                           batch size to use on training
   --no-finetune         Train network from scratch
   --optimizer OPTIMIZER
                           Optimizer to use
   --val                 Use validation on training

Pretrained checkpoint available in huggingface.

Inference KCAPTCHA with below command

usage: inference.py [-h] [--model MODEL] [--input INPUT] [--label LABEL]

options:
-h, --help     show this help message and exit
--model MODEL  Checkpoint path to use
--input INPUT  KCAPTCHA image file to read
--label LABEL  classID-to-Label mapped yaml path

On test, Model's benchmark accuracy is 97.6%

Export model to onnx with below command

python3 export_to_onnx.py --input <model path>

1. Prepare kcaptcha dataset(for best, prepare different dataset that doesn't used on training)
2. Export Quantized model with below command

   usage: export_to_onnx.py [-h] --input INPUT [--output OUTPUT] [--quantize] [--calibration-dataset CALIBRATION_DATASET] [--quant_format {QOperator,QDQ}]
                           [--per_channel PER_CHANNEL]
   
   options:
   -h, --help            show this help message and exit
   --input INPUT         input yolov8 torch model
   --output OUTPUT       output onnx model path for quantized model
   --quantize            flag for quantize onnx model. Require --calibration-dataset and --quant-format
   --calibration-dataset CALIBRATION_DATASET calibration data set path
   --quant_format {QOperator,QDQ}
   --per_channel PER_CHANNEL

Inference KCAPTCHA with below command with onnx weight

usage: inference_onnx.py [-h] [--model MODEL] [--input INPUT] [--label LABEL] [--c_thres C_THRES] [--iou_thres IOU_THRES]

options:
-h, --help            show this help message and exit
--model MODEL         Checkpoint path to use
--input INPUT         KCAPTCHA image file to read
--label LABEL         classID-to-Label mapped yaml path
--c_thres C_THRES     Confidence threshold for filtering detections.
--iou_thres IOU_THRES IoU threshold for non-maximum suppression.      

Pretrained checkpoint available in huggingface.

Only ONNX format weight accepted by server for performance reason. Quantized model(best_int8.onnx) is highly recommanded.

default exposed port is 8000

1. Install requirements with below command

   pip3 install -r requirements-server.txt

2. Start inference server with below command.

   usage: main.py [-h] [--host HOST] [--port PORT] [--model MODEL] [--label LABEL] [--c_thres C_THRES] [--iou_thres IOU_THRES]
   
   options:
   -h, --help            show this help message and exit
   --host HOST           Server host
   --port PORT           Server port
   --model MODEL         Checkpoint path to use
   --label LABEL         classID-to-Label mapped yaml path
   --c_thres C_THRES     Confidence threshold for filtering detections.
   --iou_thres IOU_THRES IoU threshold for non-maximum suppression.

docker run -d --restart always \
  -e MODEL_FILE_NAME=best_int8.onnx \
  -p 8000:8000 \
  -v ~/.model:/model \
  ghcr.io/returntofirst/kcaptcha-solver:latest

deploy model file on ~/.model and run this command.

Use docker-compose.yaml to deploy with docker-compose. deploy model file on ./model and run this command.

Use k8s-example.yaml to deploy with kubectl.