Shaun ZhouandClaude Opus 5
474b7b50a0
训练数据已经是固定边长的片，推理还把整图交给 ultralytics 就会被 letterbox 缩放， 与模型见过的形态不符。三个调用方（可视化、批量推送预测、ML backend）统一改成 切片推理再拼回原图。 - engine/tiling.merge_tiles / tiled_predict：映射回原图、裁掉镜像填充区的框、跨片去重。 推理框架留在调用方，几何部分可单测。 - engine/detection.nms：标准 NMS，另加按包含比例抑制。缺陷被切割线切开时，相邻片 只看到一角，给出的小残框与完整框 IoU 很小（面积悬殊），光靠 IoU 压不掉。 - engine/tiled_inference：ultralytics 的接线，内部分批，避免片数多时 OOM。 - 三处调用点加 --tile-size / --tile-overlap（后端为 YOLO_TILE_SIZE / YOLO_TILE_OVERLAP）， 默认 160 / 32；设 0 退回整图推理，仅用于和旧模型对比。 Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
290 lines
11 KiB
Python
290 lines
11 KiB
Python
"""固定边长切片：只裁剪与镜像填充，不缩放。"""
|
|
|
|
import pytest
|
|
|
|
np = pytest.importorskip("numpy")
|
|
|
|
from engine.detection import Box, from_yolo # noqa: E402
|
|
from engine.tiling import axis_starts, cut, plan_tiles, remap_boxes, to_source # noqa: E402
|
|
|
|
SIZE = 160
|
|
|
|
|
|
# ---------------- 切片网格 ----------------
|
|
|
|
|
|
@pytest.mark.parametrize(
|
|
"length,overlap,expected",
|
|
[
|
|
(160, 0, [0]), # 正好一片
|
|
(128, 0, [0]), # 不足，填充
|
|
(320, 0, [0, 160]), # 整两片
|
|
(200, 0, [0, 40]), # 最后一片贴边，与前片重叠 120
|
|
(480, 32, [0, 128, 256, 320]), # 步长 128，末片贴边
|
|
],
|
|
)
|
|
def test_axis_starts(length, overlap, expected):
|
|
assert axis_starts(length, SIZE, overlap) == expected
|
|
|
|
|
|
def test_last_tile_always_reaches_the_edge():
|
|
"""末尾那条不足一步的区域不能被漏掉。"""
|
|
for length in range(161, 700):
|
|
starts = axis_starts(length, SIZE, 16)
|
|
assert starts[-1] + SIZE == length
|
|
|
|
|
|
def test_plan_tiles_covers_a_large_image():
|
|
tiles = plan_tiles(524, 959, SIZE, overlap=0)
|
|
assert len(tiles) == len(axis_starts(524, SIZE, 0)) * len(axis_starts(959, SIZE, 0))
|
|
assert all(t.pad_left == 0 and t.pad_top == 0 for t in tiles) # 大图纯裁剪
|
|
|
|
|
|
def test_small_image_is_centred_and_padded():
|
|
tile, = plan_tiles(116, 83, SIZE)
|
|
assert (tile.crop_width, tile.crop_height) == (116, 83)
|
|
assert tile.pad_left == (160 - 116) // 2 == 22
|
|
assert tile.pad_top == (160 - 83) // 2 == 38
|
|
|
|
|
|
@pytest.mark.parametrize("bad", [0, -1])
|
|
def test_rejects_invalid_size(bad):
|
|
with pytest.raises(ValueError, match="size"):
|
|
plan_tiles(100, 100, bad)
|
|
|
|
|
|
@pytest.mark.parametrize("bad", [-1, 160, 200])
|
|
def test_rejects_invalid_overlap(bad):
|
|
with pytest.raises(ValueError, match="overlap"):
|
|
plan_tiles(300, 300, SIZE, bad)
|
|
|
|
|
|
# ---------------- 取片 ----------------
|
|
|
|
|
|
def test_cut_crops_without_resizing():
|
|
"""大图取片必须是原像素的子块，一个插值都不能有。"""
|
|
image = np.arange(300 * 400, dtype=np.uint8).reshape(300, 400) % 251
|
|
tile = plan_tiles(400, 300, SIZE, overlap=0)[1] # 第二片，x 有偏移
|
|
patch = cut(image, tile, SIZE)
|
|
|
|
assert patch.shape == (SIZE, SIZE)
|
|
expected = image[tile.y:tile.y + SIZE, tile.x:tile.x + SIZE]
|
|
assert np.array_equal(patch, expected)
|
|
|
|
|
|
def test_cut_mirrors_small_images():
|
|
image = np.arange(1, 5, dtype=np.uint8).reshape(1, 4) # 1x4
|
|
tile, = plan_tiles(4, 1, 4)
|
|
patch = cut(image, tile, 4)
|
|
assert patch.shape == (4, 4)
|
|
# 竖直方向镜像：原始那一行在中间，上下是它的倒影（reflect 不重复边缘行）
|
|
assert np.array_equal(patch[tile.pad_top], image[0])
|
|
assert patch.dtype == image.dtype
|
|
|
|
|
|
def test_cut_keeps_channels():
|
|
image = np.zeros((50, 60, 3), dtype=np.uint8)
|
|
tile, = plan_tiles(60, 50, SIZE)
|
|
assert cut(image, tile, SIZE).shape == (SIZE, SIZE, 3)
|
|
|
|
|
|
def test_cut_of_an_exact_size_image_is_the_image_itself():
|
|
image = np.random.default_rng(0).integers(0, 255, (SIZE, SIZE), dtype=np.uint8)
|
|
tile, = plan_tiles(SIZE, SIZE, SIZE)
|
|
assert np.array_equal(cut(image, tile, SIZE), image)
|
|
|
|
|
|
# ---------------- 框的重映射 ----------------
|
|
|
|
|
|
def test_box_shifts_with_padding():
|
|
"""小图居中填充后，框只是整体平移。"""
|
|
width, height = 80, 80
|
|
box = from_yolo(3, 0.5, 0.5, 0.25, 0.25) # 原图中心 20x20 的框
|
|
tile, = plan_tiles(width, height, SIZE)
|
|
mapped, = remap_boxes([box], tile, width, height, SIZE)
|
|
|
|
assert mapped.class_id == 3
|
|
# 原图中心落在片中心
|
|
assert (mapped.x1 + mapped.x2) / 2 == pytest.approx(0.5)
|
|
# 尺寸不变：原来 20 像素，片上还是 20 像素（占 160 的 1/8）
|
|
assert (mapped.x2 - mapped.x1) * SIZE == pytest.approx(20)
|
|
|
|
|
|
def test_box_outside_the_tile_is_dropped():
|
|
width = height = 400
|
|
box = from_yolo(0, 0.9, 0.9, 0.05, 0.05) # 右下角
|
|
first = plan_tiles(width, height, SIZE, overlap=0)[0] # 左上片
|
|
assert remap_boxes([box], first, width, height, SIZE) == []
|
|
|
|
|
|
def test_partially_visible_box_is_clipped_or_dropped():
|
|
"""跨片的框：露得多就裁剪保留，露得少就丢弃。"""
|
|
width = height = 320
|
|
# 框中心正好在 x=160 的切割线上，左右各露一半
|
|
box = from_yolo(1, 160 / width, 0.25, 40 / width, 40 / height)
|
|
left, right = plan_tiles(width, height, SIZE, overlap=0)[:2]
|
|
|
|
kept_left, = remap_boxes([box], left, width, height, SIZE, min_visibility=0.4)
|
|
assert kept_left.x2 == pytest.approx(1.0) # 裁到片的右边缘
|
|
assert (kept_left.x2 - kept_left.x1) * SIZE == pytest.approx(20)
|
|
|
|
# 阈值提到 0.6 时，只露一半就不够了
|
|
assert remap_boxes([box], left, width, height, SIZE, min_visibility=0.6) == []
|
|
assert remap_boxes([box], right, width, height, SIZE, min_visibility=0.6) == []
|
|
|
|
|
|
def test_mirrored_region_gets_no_box():
|
|
"""镜像出来的是复制品，不能再给它标框——否则等于告诉模型那里真有缺陷。"""
|
|
width = height = 80
|
|
box = from_yolo(0, 0.5, 0.5, 0.5, 0.5)
|
|
tile, = plan_tiles(width, height, SIZE)
|
|
mapped = remap_boxes([box], tile, width, height, SIZE)
|
|
|
|
assert len(mapped) == 1 # 只有一个，不是四个倒影
|
|
# 框整体落在原图内容的范围内
|
|
assert mapped[0].x1 * SIZE >= tile.pad_left - 1e-6
|
|
assert mapped[0].x2 * SIZE <= tile.pad_left + tile.crop_width + 1e-6
|
|
|
|
|
|
def test_zero_area_box_is_ignored():
|
|
tile, = plan_tiles(100, 100, SIZE)
|
|
degenerate = Box(0, 0.5, 0.5, 0.5, 0.5)
|
|
assert remap_boxes([degenerate], tile, 100, 100, SIZE) == []
|
|
|
|
|
|
# ---------------- 反向映射 ----------------
|
|
|
|
|
|
@pytest.mark.parametrize("size_wh", [(80, 60), (400, 300), (160, 160), (524, 959)])
|
|
def test_round_trip_back_to_source(size_wh):
|
|
"""片上的框映射回原图，必须落回原来的像素位置。
|
|
|
|
框放在某一片裁剪区域的正中，确保它完整落在该片内——无重叠切片时，
|
|
原图正中的框反而可能横跨切割线，不适合用来做往返校验。
|
|
"""
|
|
width, height = size_wh
|
|
for tile in plan_tiles(width, height, SIZE, overlap=0):
|
|
cx = tile.x + tile.crop_width / 2
|
|
cy = tile.y + tile.crop_height / 2
|
|
half_w = min(tile.crop_width, 20) / 4
|
|
half_h = min(tile.crop_height, 20) / 4
|
|
px = [cx - half_w, cy - half_h, cx + half_w, cy + half_h]
|
|
box = Box(2, px[0] / width, px[1] / height, px[2] / width, px[3] / height)
|
|
|
|
mapped, = remap_boxes([box], tile, width, height, SIZE, min_visibility=0.99)
|
|
back = to_source(mapped, tile, SIZE)
|
|
assert [back.x1, back.y1, back.x2, back.y2] == pytest.approx(px, abs=1e-6)
|
|
|
|
|
|
def test_score_survives_both_directions():
|
|
tile, = plan_tiles(100, 100, SIZE)
|
|
box = Box(1, 0.4, 0.4, 0.6, 0.6, 0.83)
|
|
mapped, = remap_boxes([box], tile, 100, 100, SIZE)
|
|
assert mapped.score == 0.83
|
|
assert to_source(mapped, tile, SIZE).score == 0.83
|
|
|
|
|
|
# ---------------- 切片推理的拼回 ----------------
|
|
|
|
|
|
def test_merge_deduplicates_across_the_seam():
|
|
"""重叠区里同一处缺陷会被两片各检出一次，合并后应只剩一个。"""
|
|
from engine.tiling import merge_tiles
|
|
|
|
width = height = 300
|
|
tiles = plan_tiles(width, height, SIZE, overlap=64)
|
|
# 让每片都"检出"同一个真实位置的框（像素 150..170）
|
|
per_tile = []
|
|
for tile in tiles:
|
|
px1, py1, px2, py2 = 150.0, 150.0, 170.0, 170.0
|
|
bx1 = (px1 - tile.x + tile.pad_left) / SIZE
|
|
by1 = (py1 - tile.y + tile.pad_top) / SIZE
|
|
bx2 = (px2 - tile.x + tile.pad_left) / SIZE
|
|
by2 = (py2 - tile.y + tile.pad_top) / SIZE
|
|
inside = 0 <= bx1 and bx2 <= 1 and 0 <= by1 and by2 <= 1
|
|
per_tile.append([Box(0, bx1, by1, bx2, by2, 0.9)] if inside else [])
|
|
|
|
assert sum(len(b) for b in per_tile) >= 2 # 确实有多片检出
|
|
merged = merge_tiles(per_tile, tiles, SIZE, width, height)
|
|
assert len(merged) == 1
|
|
assert merged[0].x1 * width == pytest.approx(150, abs=1e-6)
|
|
assert merged[0].x2 * width == pytest.approx(170, abs=1e-6)
|
|
|
|
|
|
def test_merge_keeps_genuinely_separate_defects():
|
|
from engine.tiling import merge_tiles
|
|
|
|
width = height = 320
|
|
tiles = plan_tiles(width, height, SIZE, overlap=0)
|
|
per_tile = [[] for _ in tiles]
|
|
per_tile[0] = [Box(0, 0.1, 0.1, 0.2, 0.2, 0.9)] # 左上片
|
|
per_tile[3] = [Box(0, 0.8, 0.8, 0.9, 0.9, 0.8)] # 右下片
|
|
assert len(merge_tiles(per_tile, tiles, SIZE, width, height)) == 2
|
|
|
|
|
|
def test_merge_clips_boxes_that_reach_into_mirrored_padding():
|
|
"""小图填充后，模型可能在镜像区出框；那里不是真实区域，要裁掉。"""
|
|
from engine.tiling import merge_tiles
|
|
|
|
width = height = 80
|
|
tiles = plan_tiles(width, height, SIZE)
|
|
# 一个横跨整片的框，两端都落在镜像填充里
|
|
merged = merge_tiles([[Box(0, 0.0, 0.4, 1.0, 0.6, 0.9)]], tiles, SIZE, width, height)
|
|
|
|
assert len(merged) == 1
|
|
assert merged[0].x1 == pytest.approx(0.0)
|
|
assert merged[0].x2 == pytest.approx(1.0) # 裁到原图边界，不超出
|
|
assert 0.0 <= merged[0].y1 <= merged[0].y2 <= 1.0
|
|
|
|
|
|
def test_merge_drops_boxes_entirely_inside_the_mirror():
|
|
from engine.tiling import merge_tiles
|
|
|
|
width = height = 80
|
|
tiles = plan_tiles(width, height, SIZE) # pad 40，原图占 40..120
|
|
ghost = Box(0, 0.05, 0.05, 0.2, 0.2, 0.9) # 像素 8..32，全在镜像区
|
|
assert merge_tiles([[ghost]], tiles, SIZE, width, height) == []
|
|
|
|
|
|
def test_tiled_predict_wires_geometry_to_the_model():
|
|
"""端到端：切片 -> 逐片推理 -> 拼回，位置要落在真实缺陷上。"""
|
|
from engine.tiling import tiled_predict
|
|
|
|
image = np.zeros((300, 400), dtype=np.uint8)
|
|
image[200:220, 100:130] = 255 # 一块亮斑
|
|
|
|
def predict_batch(patches):
|
|
"""假模型：把片里最亮的连通区域当成检出。"""
|
|
out = []
|
|
for patch in patches:
|
|
ys, xs = np.nonzero(patch == 255)
|
|
if len(xs) == 0:
|
|
out.append([])
|
|
continue
|
|
out.append([Box(0, xs.min() / SIZE, ys.min() / SIZE,
|
|
(xs.max() + 1) / SIZE, (ys.max() + 1) / SIZE, 0.9)])
|
|
return out
|
|
|
|
merged = tiled_predict(predict_batch, image, SIZE, overlap=32)
|
|
assert len(merged) == 1
|
|
assert merged[0].x1 * 400 == pytest.approx(100, abs=1)
|
|
assert merged[0].y1 * 300 == pytest.approx(200, abs=1)
|
|
assert merged[0].x2 * 400 == pytest.approx(130, abs=1)
|
|
assert merged[0].y2 * 300 == pytest.approx(220, abs=1)
|
|
|
|
|
|
def test_tiled_predict_on_a_small_image_uses_one_tile():
|
|
from engine.tiling import tiled_predict
|
|
|
|
seen = []
|
|
|
|
def predict_batch(patches):
|
|
seen.append(len(patches))
|
|
assert all(p.shape[:2] == (SIZE, SIZE) for p in patches)
|
|
return [[] for _ in patches]
|
|
|
|
tiled_predict(predict_batch, np.zeros((83, 116), dtype=np.uint8), SIZE)
|
|
assert seen == [1]
|