Signature_detectuin_custom_yolo5

Objective

We wanna detect the signature from different documents by tunning yoloy v5

Dataset

You can find it here This dataset consists of four folders:

• ‘TrainImages’ folder : contains the training images (660 images) for the detection task

• ‘TrainGroundTruth’ folder: contains the corresponding detection labels (660 files) for training detection task

• Each image in ‘TrainImages’ folder has a corresponding text file in this folder with the same name as the image

• The text file has (1 or more) rows. Each row represents the bounding box of a single signature.

• Each row has 4 values: x1,y1,x2,y2

• ‘TestImages’ folder : contains the training images (115 images) for the detection task

• ‘TestGroundTruth’ folder: contains the corresponding detection labels (115 files) for testing the detection task

Steps

1: data preparation

we have to make txt files in yolo txt file format as :

• 1: IN OUR DATASET WE HAVE DIM OF TWO POINTS INSTEAD OF center POINT , WIDTH AND HEIGHT SO WE HAVE TO CONVERT IT and normalize the dim

• 2: txt files should be in the same folder of images but in our data they aren't

I handeled them using two fuctions (convert_xy_to_xywh , data_preparing)

2: Tuning the model

• change the yaml file to contain your specific data :

  train :  /content/dataset/TrainImages
  
  val :  /content/dataset/TestImages
  
  nc : 1
  
  names : ["sinature"]
  

• run this line !python /content/yolov5/train.py --img-size 640 --batch-size 16 --epochs 5 --data dataset.yaml --cfg /content/yolov5/models/yolov5s.yaml --weights yolov5s.pt

FINALLY _Evaluation

• After getting your CUSTOMIZED YOLO, you can find it in this path ( yolov5/runs/train/exp/weights/last.pt)

• test it at any image