cascadeCNN_license_plate_detection

Implement cascade cnn for license plate detection

Author: HuanQin

E-mail: xiaoyu1qh1@163.com

Paper : http://users.eecs.northwestern.edu/~xsh835/assets/cvpr2015_cascnn.pdf

Train process

Train process details in process.txt

• preprocess_data : create positive data and negative data, resize, write file list, test recall
• lmdb : change data format to lmdb
• train_net : train net
• script : no use

Test process

Test process details in lp_test.py, you can run python lp_test.py

You need to change some parameters as follows:

• caffe_root : caffe root dir
• workspace : code dir
• img_dir : image dir
• img_list_file: image list file
• min_lp_size : minimum license plate height size
• max_lp_size : maximum license plate height size
• save_res_dir : save result dir

run lp_test.py

• load model
• detect license plate
• save results

I set up the ratio of w and h to 3:1. net input size is as follow:

• 12-net : 12x4
• 12-cal : 36x12
• 24-net : 36x12
• 24-cal : 36x12
• 48-net : 72x24
• 48-cal : 72x24

For my dataset, I only use 12-net, 12-cal-net, 24-net and 48-cal-net.

You can change the parameters if you want.

More information, you can read the paper and see the code.

results

Use 12-net, 12-cal-net, 24-net and 48-cal-net, runs at 10 FPS on a single CPU(Intel(R) Xeon(R) CPU E5-2630 v4 @ 2.20GHz) for 640x360 images.

For more accurary, you can use 12-net, 12-cal, 24-net, 24-cal, 48-net and 48-cal.

Detection results: