rotate-captcha-crack

lumina37/rotate-captcha-crack is ranked 18,262 among these 8 peers; PengBoXiangShang/SiamRPN_plus_plus_PyTorch is ranked 18,253.

This project can be used for rotate-captcha cracking. Test result: Three kinds of model are implemented, as shown in the table below. 18G | RotNet is the implementation of over PyTorch. is based on , with as its backbone and class number of 128. The average prediction error is , obtained by 128 epochs of training (3. 4 hours) on the COCO 2017 (Unlabeled) Dataset. The Cross-Domain Test uses COCO 2017 (Unlabeled) Dataset for training, and Captcha Pictures from Baidu (thanks

to @xiangbei1997) for testing. The captcha picture used in the demo above comes from RotateCaptchaBreak Try it! Prepare CUDA device with mem>=16G for training (reduce the batch size if necessary) Python>=3. This project strongly suggest you to use for package management. Run the following commands if you already have : Or, if you prefer : The following steps will create a virtual env under the working directory. Or, if you prefer a direct : Download the Pretrained Models Download the files

in Release and unzip them all to the dir. The directory structure will be like The names of models will change frequently as the project is still in beta status. So, if any occurs, please try to rollback to the corresponding tag first. Test the Rotation Effect by a Single Captcha Picture Open to check the result. If you do not have , please use: Use HTTP Server Install extra dependencies With : or with : or with : Launch server If you do

not have , just use: Another Shell to Send Images Use curl: Or use Windows PowerShell: Train Your Own Model Prepare Datasets For this project I'm using Google Street View and Landscape-Dataset for training. You can collect some photos and leave them in one directory. Modify the variable in , let it points to the directory containing images. All the cropping, rotation and resizing will be done soon after the image is loaded. Train Validate the Model on Test Set Details

of Design Most of the rotate-captcha cracking methods are based on , with as its backbone. regards the angle prediction as a classification task with 360 classes, then uses cross entropy to compute the loss. Yet cross-entropy loss with one-hot labels will bring a uniform metric distance between all angles (e. $\mathrm{dist}(1°, 2°) = \mathrm{dist}(1°, 180°)$ ), clearly defies the common sense. *Arbitrary-Oriented Object Detection with Circular Smooth Label (ECCV'20)* introduces an interesting trick

, by smoothing the one-hot label, e. , CSL provides a loss measurement closer to our intuition, such that $\mathrm{dist}(1°,180°) \gt \mathrm{dist}(1°,3°)$. Meanwhile, the proposed by d4nst/RotNet is less effective. That's because when dealing with outliers, the gradient leads to a non-convergence result.