Implement cascade cnn for license plate detection
E-mail: xiaoyu1qh1@163.com
Train process details in
process.txt
preprocess_data
: create positive data and negative data, resize, write file list, test recalllmdb
: change data format to lmdbtrain_net
: train netscript
: no use
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 dirworkspace
: code dirimg_dir
: image dirimg_list_file
: image list filemin_lp_size
: minimum license plate height sizemax_lp_size
: maximum license plate height sizesave_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
: 12x412-cal
: 36x1224-net
: 36x1224-cal
: 36x1248-net
: 72x2448-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.
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: