feat: enhance DOCX processing by refining image handling and improving logging for inference timing

demo/demo.py

@@ -218,7 +218,7 @@ def _process_output(
         make_func = pipeline_union_make
     elif process_mode == "vlm":
         make_func = vlm_union_make
-    elif process_mode == "office":
+    elif process_mode in office_suffixes:
         make_func = office_union_make
     else:
         raise Exception(f"Unknown process_mode: {process_mode}")

mineru/backend/office/docx_analyze.py

@@ -17,12 +17,13 @@ def office_docx_analyze(
     results = convert_binary(file_stream)
 
     infer_time = round(time.time() - infer_start, 2)
-    logger.debug(f"infer finished, cost: {infer_time}, speed: {round(len(results) / infer_time, 3)} page/s")
+    safe_time = max(infer_time, 0.01)
+    logger.debug(f"infer finished, cost: {infer_time}, speed: {round(len(results) / safe_time, 3)} page/s")
 
     # middle_json = result_to_middle_json(
     #     results,
     #     image_writer,
     # )
-    middle_json= []
+    middle_json= {"pdf_info": results}
 
     return middle_json, results

mineru/model/docx/docx_converter.py

@@ -5,7 +5,7 @@ from pathlib import Path
 from typing import BinaryIO, Optional, Union, Any, Final
 
 import logging
-from PIL import Image
+from PIL import Image, WmfImagePlugin
 from loguru import logger
 from docx import Document
 from docx.oxml.xmlchemy import BaseOxmlElement
@@ -466,7 +466,14 @@ class DocxConverter:
         else:
             image_bytes = BytesIO(image_data)
             pil_image = Image.open(image_bytes)
-            img_base64 = image_to_b64str(pil_image)
+            if isinstance(pil_image, WmfImagePlugin.WmfStubImageFile):
+                logger.warning(f"Skipping WMF image, size: {pil_image.size}")
+                placeholder = Image.new('RGB', pil_image.size, (240, 240, 240))
+                img_base64 = image_to_b64str(placeholder)
+            else:
+                if pil_image.mode != "RGB":
+                    pil_image = pil_image.convert("RGB")
+                img_base64 = image_to_b64str(pil_image)
             image_block = {
                 "type": BlockType.IMAGE,
                 "bbox": [0, 0, 0, 0],