[Feature] 是否可以添加自定义图像识别模型

[LyricaPrismriver]

Description

自带的yolo只能转文字，对于一些游戏图标的识别效果很差，很多游戏按钮都是只有图标没有文字的，agent发现文字内容没有想要的东西就切图像识别模型识别不容易卡死

Use Case

多模态模型辅助识别纯图标游戏按钮

Implementation Ideas

No response

Additional Information

No response

[alectimison-maker]

I inspected the current main branch (f809ff1) and would be interested in contributing a bounded first slice, if this direction matches the maintainers' intent.

Today UIParser constructs the ONNX session directly and _detect_icons() assumes the bundled YOLO output shape before emitting generic icon labels. That makes an alternative icon recognizer or VLM fallback difficult to add without changing the OCR/merge pipeline at the same time.

My proposed first PR would:

• introduce a small icon-detector boundary that returns the existing normalized box shape;
• preserve the current YOLO/ONNX detector as the default with no configuration or behavior change;
• allow a custom detector to supply semantic labels while reusing the existing OCR overlap and coordinate normalization;
• fail open to OCR-only parsing when a custom detector fails;
• add unit tests for default compatibility, custom semantic labels, merge behavior, and detector failures.

I would deliberately not add a new hosted model dependency or make a VLM call on every frame in this first PR. Those choices affect latency, cost, privacy, and configuration, while the detector boundary is useful for either a local model or a later opt-in VLM fallback.

Would this be an acceptable first contribution for this issue, or would you prefer that implementation wait until an end-to-end VLM/configuration design is agreed?