Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

feat(add): Special Image Classifier based on YOLOv5 #15

Closed
QIN2DIM opened this issue Jan 28, 2022 · 1 comment
Closed

feat(add): Special Image Classifier based on YOLOv5 #15

QIN2DIM opened this issue Jan 28, 2022 · 1 comment
Assignees
Labels
feature 新特性或新需求 fixed BUG 已修复或问题已解决

Comments

@QIN2DIM
Copy link
Owner

QIN2DIM commented Jan 28, 2022

不过这条特性大概率是不会用上的,本项目的需求只是领个游戏而已。 不行我就要一次过!

在保证可玩性的情况下,针对地训练一个最小化的 YOLOv5 图像分类器(杀鸡用核弹)。

目前一轮挑战耗时 12 秒左右(存在无法覆盖的类别),针对性训练后的识别效率可以提升 3 至 10 倍,预计耗时压榨至 3 秒内,极端情况下可在模型体积更小,占用内存更低的情况下,秒解挑战。

  • 感兴趣的朋友可以参考这篇论文自己玩一玩 10.1109/SPW53761.2021.00061
  • 根据文章中提到的“技术建议”可知,hCaptcha challenge 的图像类型数量在短时间内不会有太大变化(从文章发表至今删去了 雨伞 新添了 巴土水上飞机),为此,我们可以刻意地训练一个完全适配度的模型用来处理人机挑战(也即所有可供分类的图像数量较少,正常作业中不会遇到突发情况,相当于模拟考前做了一遍原题还对了参考答案,即使遇见突发情况也可以借助强大的关联点网络做排除法)。
@QIN2DIM QIN2DIM added the feature 新特性或新需求 label Jan 28, 2022
@QIN2DIM QIN2DIM changed the title feat(preview): Special Image Classifier based on YOLOv5-mini feat(add): Special Image Classifier based on YOLOv5-mini Feb 4, 2022
@QIN2DIM QIN2DIM changed the title feat(add): Special Image Classifier based on YOLOv5-mini feat(preview): Special Image Classifier based on YOLOv5-mini Feb 9, 2022
@QIN2DIM QIN2DIM changed the title feat(preview): Special Image Classifier based on YOLOv5-mini feat(add): Special Image Classifier based on YOLOv5m6 Feb 13, 2022
@QIN2DIM QIN2DIM changed the title feat(add): Special Image Classifier based on YOLOv5m6 feat(add): Special Image Classifier based on YOLOv5 Feb 13, 2022
@QIN2DIM
Copy link
Owner Author

QIN2DIM commented Feb 14, 2022

闪电解算

使用 YOLOv5m6(onnx) 执行 claim 指令的运行日志如下图所示:

Snipaste_2022-02-14_01-16-35

各方案对比

下表统计了不同规模的 onnx model 处理 30 轮(每 9 张挑战图片记数一轮)混合类别的 hCAPTCHA 人机挑战的平均解算耗时:

model(onnx) avg_time(s) model_size
yolov5n6 0.71 12.4MB
yolov5s6 1.422 48.2MB
yolov5m6 3.05 136MB

使用 YOLOv5n6(onnx) 迁移模型作为临时替代方案,获得接近极限的解算速度;使用 YOLOv5s6(onnx) 可以在稳定性,功耗,解算效率之间保持最佳平衡。

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
feature 新特性或新需求 fixed BUG 已修复或问题已解决
Projects
Development

No branches or pull requests

1 participant