升级 PaddleOCR 至 3.4.0，移除 PP-OCRv4 模型

- PaddleOCR 2.10.0 升级至 3.4.0，使用新的 TextDetection API 替代旧版 TextDetector
- V5 模型替换为 PaddleX 格式（inference.pdiparams + inference.yml）
- 移除 PP-OCRv4 模型（backend/models/V4），仅保留 PP-OCRv5
- 字幕检测选项重命名为"快速"和"精准"
- 清理 requirements.txt，移除 11 个未使用的依赖，补充缺失的依赖
- 更新设计素材

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

backend/config.py

@@ -52,7 +52,7 @@ class Config(QConfig):
     # 【设置inpaint算法】
     inpaintMode = OptionsConfigItem("Main", "InpaintMode", InpaintMode.STTN_AUTO, OptionsValidator(InpaintMode), EnumSerializer(InpaintMode))
     
-    subtitleDetectMode =  OptionsConfigItem("Main", "SubtitleDetectMode", SubtitleDetectMode.PP_OCRv4_SERVER, OptionsValidator(SubtitleDetectMode), EnumSerializer(SubtitleDetectMode))
+    subtitleDetectMode =  OptionsConfigItem("Main", "SubtitleDetectMode", SubtitleDetectMode.PP_OCRv5_SERVER, OptionsValidator(SubtitleDetectMode), EnumSerializer(SubtitleDetectMode))
 
     # 【设置像素点偏差】
     # 用于判断是不是非字幕区域(一般认为字幕文本框的长度是要大于宽度的，如果字幕框的高大于宽，且大于的幅度超过指定像素点大小，则认为是错误检测)

backend/models/V4/ch_det/fs_manifest.csv

@@ -1,4 +0,0 @@
-filename,filesize,encoding,header␍
-inference_1.onnx,50000000,,␍
-inference_2.onnx,50000000,,␍
-inference_3.onnx,13443410,,␍

backend/models/V5/ch_det/config.json

@@ -0,0 +1,111 @@
+{
+    "Global": {
+        "model_name": "PP-OCRv5_server_det"
+    },
+    "Hpi": {
+        "backend_configs": {
+            "paddle_infer": {
+                "trt_dynamic_shapes": {
+                    "x": [
+                        [
+                            1,
+                            3,
+                            32,
+                            32
+                        ],
+                        [
+                            1,
+                            3,
+                            736,
+                            736
+                        ],
+                        [
+                            1,
+                            3,
+                            4000,
+                            4000
+                        ]
+                    ]
+                }
+            },
+            "tensorrt": {
+                "dynamic_shapes": {
+                    "x": [
+                        [
+                            1,
+                            3,
+                            32,
+                            32
+                        ],
+                        [
+                            1,
+                            3,
+                            736,
+                            736
+                        ],
+                        [
+                            1,
+                            3,
+                            4000,
+                            4000
+                        ]
+                    ]
+                }
+            }
+        }
+    },
+    "PreProcess": {
+        "transform_ops": [
+            {
+                "DecodeImage": {
+                    "channel_first": false,
+                    "img_mode": "BGR"
+                }
+            },
+            {
+                "DetLabelEncode": null
+            },
+            {
+                "DetResizeForTest": {
+                    "resize_long": 960
+                }
+            },
+            {
+                "NormalizeImage": {
+                    "mean": [
+                        0.485,
+                        0.456,
+                        0.406
+                    ],
+                    "order": "hwc",
+                    "scale": "1./255.",
+                    "std": [
+                        0.229,
+                        0.224,
+                        0.225
+                    ]
+                }
+            },
+            {
+                "ToCHWImage": null
+            },
+            {
+                "KeepKeys": {
+                    "keep_keys": [
+                        "image",
+                        "shape",
+                        "polys",
+                        "ignore_tags"
+                    ]
+                }
+            }
+        ]
+    },
+    "PostProcess": {
+        "name": "DBPostProcess",
+        "thresh": 0.3,
+        "box_thresh": 0.6,
+        "max_candidates": 1000,
+        "unclip_ratio": 1.5
+    }
+}

backend/models/V5/ch_det/fs_manifest.csv

@@ -1,3 +0,0 @@
-filename,filesize,encoding,header␍
-inference_1.onnx,50000000,,␍
-inference_2.onnx,38118768,,␍

backend/models/V5/ch_det/inference.yml

@@ -0,0 +1,53 @@
+Global:
+  model_name: PP-OCRv5_server_det
+Hpi:
+  backend_configs:
+    paddle_infer:
+      trt_dynamic_shapes: &id001
+        x:
+        - - 1
+          - 3
+          - 32
+          - 32
+        - - 1
+          - 3
+          - 736
+          - 736
+        - - 1
+          - 3
+          - 4000
+          - 4000
+    tensorrt:
+      dynamic_shapes: *id001
+PreProcess:
+  transform_ops:
+  - DecodeImage:
+      channel_first: false
+      img_mode: BGR
+  - DetLabelEncode: null
+  - DetResizeForTest:
+      resize_long: 960
+  - NormalizeImage:
+      mean:
+      - 0.485
+      - 0.456
+      - 0.406
+      order: hwc
+      scale: 1./255.
+      std:
+      - 0.229
+      - 0.224
+      - 0.225
+  - ToCHWImage: null
+  - KeepKeys:
+      keep_keys:
+      - image
+      - shape
+      - polys
+      - ignore_tags
+PostProcess:
+  name: DBPostProcess
+  thresh: 0.3
+  box_thresh: 0.6
+  max_candidates: 1000
+  unclip_ratio: 1.5

backend/models/V5/ch_det_fast/config.json

@@ -0,0 +1,111 @@
+{
+    "Global": {
+        "model_name": "PP-OCRv5_mobile_det"
+    },
+    "Hpi": {
+        "backend_configs": {
+            "paddle_infer": {
+                "trt_dynamic_shapes": {
+                    "x": [
+                        [
+                            1,
+                            3,
+                            32,
+                            32
+                        ],
+                        [
+                            1,
+                            3,
+                            736,
+                            736
+                        ],
+                        [
+                            1,
+                            3,
+                            4000,
+                            4000
+                        ]
+                    ]
+                }
+            },
+            "tensorrt": {
+                "dynamic_shapes": {
+                    "x": [
+                        [
+                            1,
+                            3,
+                            32,
+                            32
+                        ],
+                        [
+                            1,
+                            3,
+                            736,
+                            736
+                        ],
+                        [
+                            1,
+                            3,
+                            4000,
+                            4000
+                        ]
+                    ]
+                }
+            }
+        }
+    },
+    "PreProcess": {
+        "transform_ops": [
+            {
+                "DecodeImage": {
+                    "channel_first": false,
+                    "img_mode": "BGR"
+                }
+            },
+            {
+                "DetLabelEncode": null
+            },
+            {
+                "DetResizeForTest": {
+                    "resize_long": 960
+                }
+            },
+            {
+                "NormalizeImage": {
+                    "mean": [
+                        0.485,
+                        0.456,
+                        0.406
+                    ],
+                    "order": "hwc",
+                    "scale": "1./255.",
+                    "std": [
+                        0.229,
+                        0.224,
+                        0.225
+                    ]
+                }
+            },
+            {
+                "ToCHWImage": null
+            },
+            {
+                "KeepKeys": {
+                    "keep_keys": [
+                        "image",
+                        "shape",
+                        "polys",
+                        "ignore_tags"
+                    ]
+                }
+            }
+        ]
+    },
+    "PostProcess": {
+        "name": "DBPostProcess",
+        "thresh": 0.3,
+        "box_thresh": 0.6,
+        "max_candidates": 1000,
+        "unclip_ratio": 1.5
+    }
+}

backend/models/V5/ch_det_fast/inference.yml

@@ -0,0 +1,53 @@
+Global:
+  model_name: PP-OCRv5_mobile_det
+Hpi:
+  backend_configs:
+    paddle_infer:
+      trt_dynamic_shapes: &id001
+        x:
+        - - 1
+          - 3
+          - 32
+          - 32
+        - - 1
+          - 3
+          - 736
+          - 736
+        - - 1
+          - 3
+          - 4000
+          - 4000
+    tensorrt:
+      dynamic_shapes: *id001
+PreProcess:
+  transform_ops:
+  - DecodeImage:
+      channel_first: false
+      img_mode: BGR
+  - DetLabelEncode: null
+  - DetResizeForTest:
+      resize_long: 960
+  - NormalizeImage:
+      mean:
+      - 0.485
+      - 0.456
+      - 0.406
+      order: hwc
+      scale: 1./255.
+      std:
+      - 0.229
+      - 0.224
+      - 0.225
+  - ToCHWImage: null
+  - KeepKeys:
+      keep_keys:
+      - image
+      - shape
+      - polys
+      - ignore_tags
+PostProcess:
+  name: DBPostProcess
+  thresh: 0.3
+  box_thresh: 0.6
+  max_candidates: 1000
+  unclip_ratio: 1.5

backend/tools/constant.py

@@ -16,7 +16,5 @@ class SubtitleDetectMode(Enum):
     """
     字幕检测算法枚举
     """
-    PP_OCRv4_SERVER = "PP-OCRv4-Server" 
+    PP_OCRv5_MOBILE = "快速"
-    PP_OCRv4_MOBILE = "PP-OCRv4-Mobile"
+    PP_OCRv5_SERVER = "精准"
-    PP_OCRv5_SERVER = "PP-OCRv5-Server" 
-    PP_OCRv5_MOBILE = "PP-OCRv5-Mobile" 

backend/tools/model_config.py

@@ -3,6 +3,11 @@ from backend.config import config, BASE_DIR
 from backend.tools.common_tools import merge_big_file_if_not_exists
 from backend.tools.constant import SubtitleDetectMode
 
+_MODEL_NAME_MAP = {
+    SubtitleDetectMode.PP_OCRv5_MOBILE: "PP-OCRv5_mobile_det",
+    SubtitleDetectMode.PP_OCRv5_SERVER: "PP-OCRv5_server_det",
+}
+
 class ModelConfig:
     def __init__(self):
         self.LAMA_MODEL_DIR = os.path.join(BASE_DIR, 'models', 'big-lama')
@@ -13,14 +18,9 @@ class ModelConfig:
             self.DET_MODEL_DIR = os.path.join(BASE_DIR,'models', 'V5', 'ch_det_fast')
         elif config.subtitleDetectMode.value == SubtitleDetectMode.PP_OCRv5_SERVER:
             self.DET_MODEL_DIR = os.path.join(BASE_DIR, 'models', 'V5', 'ch_det')
-        elif config.subtitleDetectMode.value == SubtitleDetectMode.PP_OCRv4_MOBILE:
-            self.DET_MODEL_DIR = os.path.join(BASE_DIR,'models', 'V4', 'ch_det_fast')
-        elif config.subtitleDetectMode.value == SubtitleDetectMode.PP_OCRv4_SERVER:
-            self.DET_MODEL_DIR = os.path.join(BASE_DIR, 'models', 'V4', 'ch_det')
         else:
             raise ValueError(f"Invalid subtitle detect mode: {config.subtitleDetectMode.value}")
+        self.DET_MODEL_NAME = _MODEL_NAME_MAP[config.subtitleDetectMode.value]
 
         merge_big_file_if_not_exists(self.LAMA_MODEL_DIR, 'bit-lama.pt')
         merge_big_file_if_not_exists(self.PROPAINTER_MODEL_DIR, 'ProPainter.pth')
-        merge_big_file_if_not_exists(self.DET_MODEL_DIR, 'inference.onnx')
-    

backend/tools/subtitle_detect.py

@@ -1,4 +1,3 @@
-import os
 import sys
 from functools import cached_property
 
@@ -27,36 +26,37 @@ class SubtitleDetect:
     def text_detector(self):
         import paddle
         paddle.disable_signal_handler()
-        from paddleocr.tools.infer import utility
+        from paddleocr import TextDetection
-        from paddleocr.tools.infer.predict_det import TextDetector
         hardware_accelerator = HardwareAccelerator.instance()
         onnx_providers = hardware_accelerator.onnx_providers
         model_config = ModelConfig()
-        parser = utility.init_args()
+        return TextDetection(
-        args = parser.parse_args([])
+            model_name=model_config.DET_MODEL_NAME,
-        args.det_algorithm = 'DB'
+            model_dir=model_config.DET_MODEL_DIR,
-        args.det_model_dir = os.path.join(model_config.DET_MODEL_DIR, 'inference.onnx') if len(onnx_providers) > 0 else model_config.DET_MODEL_DIR
+            device="cpu",
-        args.use_gpu=False
+            enable_hpi=len(onnx_providers) > 0,
-        args.use_onnx=len(onnx_providers) > 0
+        )
-        args.onnx_providers=onnx_providers
-        return TextDetector(args)
 
     def detect_subtitle(self, img):
         temp_list = []
-        dt_boxes, elapse = self.text_detector(img)
+        results = self.text_detector.predict(img)
-        coordinate_list = get_coordinates(dt_boxes.tolist())
+        for res in results:
-        if coordinate_list:
+            dt_polys = res['dt_polys']
-            for coordinate in coordinate_list:
+            if dt_polys is None or len(dt_polys) == 0:
-                xmin, xmax, ymin, ymax = coordinate
+                continue
-                if self.sub_areas is not None and len(self.sub_areas) > 0:
+            coordinate_list = get_coordinates(dt_polys.tolist())
-                    for sub_area in self.sub_areas:
+            if coordinate_list:
-                        s_ymin, s_ymax, s_xmin, s_xmax = sub_area
+                for coordinate in coordinate_list:
-                        if (s_xmin <= xmin and xmax <= s_xmax
+                    xmin, xmax, ymin, ymax = coordinate
-                                and s_ymin <= ymin
+                    if self.sub_areas is not None and len(self.sub_areas) > 0:
-                                and ymax <= s_ymax):
+                        for sub_area in self.sub_areas:
-                            temp_list.append((xmin, xmax, ymin, ymax))
+                            s_ymin, s_ymax, s_xmin, s_xmax = sub_area
-                else:
+                            if (s_xmin <= xmin and xmax <= s_xmax
-                    temp_list.append((xmin, xmax, ymin, ymax))
+                                    and s_ymin <= ymin
+                                    and ymax <= s_ymax):
+                                temp_list.append((xmin, xmax, ymin, ymax))
+                    else:
+                        temp_list.append((xmin, xmax, ymin, ymax))
         return temp_list
 
     def find_subtitle_frame_no(self, sub_remover=None):

requirements.txt

@@ -1,25 +1,28 @@
-albumentations~=1.4.10
 filesplit==3.0.2
 opencv-python==4.11.0.86
 scikit-image==0.25.2
-imgaug==0.4.0
-pyclipper==1.3.0.post6
-lmdb==1.6.2
-PyYAML==6.0.2
-omegaconf==2.3.0
 tqdm==4.67.1
 pyside6-fluent-widgets==1.7.7
-easydict==1.13
-scikit-learn==1.6.1
-pandas==2.2.3
-webdataset==0.2.111
 numpy==2.2.5
-protobuf==6.30.2
+av==17.0.0
-av==14.3.0
 einops==0.8.1
-paddleocr==2.10.0
+paddleocr==3.4.0
-onnxruntime-gpu==1.20.1; sys_platform == 'linux'
+matplotlib
+Pillow
+scipy
+requests
+darkdetect
+je-showinfilemanager==1.1.6a4
+# torch and torchvision are installed separately (see Dockerfile for platform-specific versions)
+
+# Training only
+#matplotlib
+#lpips
+#tensorboardX
+
+# Hardware acceleration
+#onnxruntime-gpu==1.20.1; sys_platform == 'linux'
+
 # do not upgrade, windows 10/11 compatable issues, check hardware_acclerator.py:100
-onnxruntime-directml==1.20.1; sys_platform == 'win32'
+#onnxruntime-directml==1.20.1; sys_platform == 'win32'
-onnxruntime==1.22.0; sys_platform == 'darwin'
+#onnxruntime==1.22.0; sys_platform == 'darwin'
-je-showinfilemanager==1.1.6a4