在 OCR 识别领域最权威的会议之一 ICDAR(国际文档分析与识别会议)上,360 数科在 ICDAR2019- SROIE 榜单上斩获第一。
![图片](https://image.jiqizhixin.com/uploads/editor/5aacb422-7ce9-4bcd-aa69-ec8fb1753f65/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/319995ea-8f47-4870-9a3b-c79cdb6bc626/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/4ef0b9f8-8862-4b37-82e8-95bc08fc9803/640.png)
文本行字体模糊不清。官方给出的比赛数据集,均来自商超结算小票扫描图像,由于小票均为机打且存放时间过长,导致扫描出来的文本行存在较为严重的磨损和缺失,字体笔画不完整等情形,这给 OCR 识别算法带来很大挑战。 文本行图像出现弯曲。给出的文本行图像中出现较大比例的弯曲,现今主流文本行识别算法对水平文本识别较为稳健,弯曲文本行识别是 OCR 识别业内难点。 标注歧义。给出来的文本行在对应的文本图像中根本不存在、空格标注错误以及形近字标注错误,这给算法的泛化性带来了很大的冲击。
![图片](https://image.jiqizhixin.com/uploads/editor/2488c205-f87f-46b8-80e0-fcdf7582810f/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/4b286e1e-ba67-4690-902c-655cbd705627/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/f1e254af-61aa-49a6-b858-cbc52c20722d/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/741b0f39-5d60-4fcd-a678-cf87dfaab32c/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/83836534-ac2a-4478-96a8-9e0d5101bb7f/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/2f68e69f-8aeb-4d24-8272-55c3e863f311/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/101c2795-5b78-4853-9cc4-200b8e5a9b9c/640.png)