my blog about CTPN blog
Ctpn works better for horizontal text, but worse for oblique and curved text.If you are interested in text detection and model compression, you can pay attention to the two projects I will do next, psenet and DBNet
nms and bbox utils are written in cython, you have to build the library first.
cd utils/bbox
sh make.shIt will generate a nms.so and a bbox.so in current folder.
- follow setup to build the library
- download the test model
- change you own model_path , dir_path and save_path in inference.py
python3 inference.py
| base_model | Model_size(M) | model_file |
|---|---|---|
| vgg16_bn | 67.7 | baiduyun(extract code: 5pgy) |
| resnet50 | 137 | baiduyun(extract code: 5pgy) |
| shufflenet_v2_x1_0 | 25.4 | baiduyun (extract code: 5pgy) |
| mobilenet_v3_large | 16.9 | baiduyun (extract code: 5pgy) |
| mobilenet_v3_small | 13.5 | baiduyun (extract code: 5pgy) |
follow icdar15 dataset format, x1,y1,x2,y2,x3,y3,x4,y4,label
image
│ 1.jpg
│ 2.jpg
│ ...
label
│ 1.txt
│ 2.txt
| ...
Simplely run
python3 train.py --base_model vgg16_bn --batch_size 4 --size_list [1048]
- Support switching basemodel,(mobilenet_v3_large,mobilenet_v3_small, shufflenet_v2_x1_0, shufflenet_v2_x0_5, vgg11, vgg11_bn, vgg16, vgg16_bn, vgg19, vgg19_bn, resnet18, resnet34 ,resnet50, resnet101, resnet152)
- Ohem algorithm is added
- Support batch training
- When the size_list has multiple values, the maximum edge of the training picture will be randomly zoomed when training. It should be noted that you must ensure that your GPU memory supports maximum edge scaling.
| rescall | prediction | hmean |
|---|---|---|
| 0.4058 | 0.6117 | 0.4879 |
