fix(det): honour limit_side_len when limit_type is max (issue #740) (#742)

TextDetector.get_preprocess replaced limit_side_len with a hard-coded
960/1500/2000 bucket whenever limit_type was "max", so the configured
value was ignored and large images reached the detector unresized.
Pass the configured limit_side_len through for both limit types, matching
PaddleOCR DetResizeForTest.

python/rapidocr/ch_ppocr_det/main.py

@@ -53,7 +53,7 @@ class TextDetector:
             raise ValueError("img is None")
 
         ori_img_shape = img.shape[0], img.shape[1]
-        self.preprocess_op = self.get_preprocess(max(img.shape[0], img.shape[1]))
+        self.preprocess_op = self.get_preprocess()
         prepro_img = self.preprocess_op(img)
         if prepro_img is None:
             return TextDetOutput()
@@ -67,16 +67,8 @@ class TextDetector:
         elapse = time.perf_counter() - start_time
         return TextDetOutput(img, boxes, scores, elapse=elapse)
 
-    def get_preprocess(self, max_wh: int) -> DetPreProcess:
-        if self.limit_type == "min":
-            limit_side_len = self.limit_side_len
-        elif max_wh < 960:
-            limit_side_len = 960
-        elif max_wh < 1500:
-            limit_side_len = 1500
-        else:
-            limit_side_len = 2000
-        return DetPreProcess(limit_side_len, self.limit_type, self.mean, self.std)
+    def get_preprocess(self) -> DetPreProcess:
+        return DetPreProcess(self.limit_side_len, self.limit_type, self.mean, self.std)
 
     @staticmethod
     def sorted_boxes(dt_boxes: np.ndarray) -> np.ndarray:

python/tests/test_det_preprocess.py

@@ -0,0 +1,43 @@
+# -*- encoding: utf-8 -*-
+# @Author: Mohammad Hijjawi
+import numpy as np
+import pytest
+from rapidocr.ch_ppocr_det import TextDetector
+
+
+def get_detector(limit_side_len: int, limit_type: str) -> TextDetector:
+    detector = TextDetector.__new__(TextDetector)
+    detector.limit_side_len = limit_side_len
+    detector.limit_type = limit_type
+    detector.mean = [0.5, 0.5, 0.5]
+    detector.std = [0.5, 0.5, 0.5]
+    detector.postprocess_op = lambda preds, ori_shape: ([], [])
+    return detector
+
+
+@pytest.mark.parametrize(
+    "limit_side_len,limit_type,img_hw,expected_hw",
+    [
+        # issue #740: limit_side_len must be honoured when limit_type is "max"
+        (960, "max", (2000, 2000), (960, 960)),
+        (960, "max", (1000, 2000), (480, 960)),
+        (640, "max", (1200, 800), (640, 416)),
+        (960, "max", (320, 640), (320, 640)),
+        (736, "min", (500, 1000), (736, 1472)),
+        (736, "min", (1000, 800), (992, 800)),
+        (736, "min", (800, 1600), (800, 1600)),
+    ],
+)
+def test_det_preprocess_respects_limit_side_len(
+    limit_side_len, limit_type, img_hw, expected_hw
+):
+    detector = get_detector(limit_side_len, limit_type)
+    img = np.zeros((*img_hw, 3), dtype=np.uint8)
+
+    det_inputs = []
+    detector.session = det_inputs.append
+
+    detector(img)
+
+    assert len(det_inputs) == 1
+    assert det_inputs[0].shape == (1, 3, *expected_hw)