Add dockerized detector and UI cleanup

app/vision.py

@@ -2,6 +2,7 @@ from __future__ import annotations
 
 from dataclasses import dataclass
 from io import BytesIO
+from pathlib import Path
 from typing import Any
 
 import cv2
@@ -24,6 +25,9 @@ class BBox:
         return self.y + self.h
 
 
+FACE_CASCADE = cv2.CascadeClassifier(
+    str(Path(cv2.data.haarcascades) / "haarcascade_frontalface_default.xml")
+)
 HOG = cv2.HOGDescriptor()
 HOG.setSVMDetector(cv2.HOGDescriptor_getDefaultPeopleDetector())
 
@@ -36,20 +40,59 @@ def decode_image(image_bytes: bytes) -> np.ndarray:
     return image
 
 
-def select_primary_bbox(image: np.ndarray) -> tuple[BBox, str]:
+def select_primary_bbox(image: np.ndarray, detector: str = "auto") -> tuple[BBox, str]:
+    detector = (detector or "auto").strip().lower()
+
+    if detector == "face":
+        face_bbox = detect_face(image)
+        if face_bbox is not None:
+            return face_bbox, "face_cascade"
+        return fallback_bbox(image), "center_fallback"
+
+    if detector == "person":
+        person_bbox = detect_person(image)
+        if person_bbox is not None:
+            return person_bbox, "person_hog"
+        return fallback_bbox(image), "center_fallback"
+
+    if detector == "salient":
+        salient_bbox = detect_salient_object(image)
+        if salient_bbox is not None:
+            return salient_bbox, "salient_contour"
+        return fallback_bbox(image), "center_fallback"
+
+    face_bbox = detect_face(image)
+    if face_bbox is not None:
+        return face_bbox, "face_cascade"
+
     person_bbox = detect_person(image)
     if person_bbox is not None:
         return person_bbox, "person_hog"
 
-    contour_bbox = detect_salient_object(image)
-    if contour_bbox is not None:
-        return contour_bbox, "contour"
+    salient_bbox = detect_salient_object(image)
+    if salient_bbox is not None:
+        return salient_bbox, "salient_contour"
 
+    return fallback_bbox(image), "center_fallback"
+
+
+def fallback_bbox(image: np.ndarray) -> BBox:
     h, w = image.shape[:2]
-    side = int(min(w, h) * 0.8)
+    side = int(min(w, h) * 0.85)
+    side = max(1, min(side, w, h))
     x = max(0, (w - side) // 2)
     y = max(0, (h - side) // 2)
-    return BBox(x=x, y=y, w=side, h=side), "center_fallback"
+    return BBox(x=x, y=y, w=side, h=side)
+
+
+def detect_face(image: np.ndarray) -> BBox | None:
+    gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
+    gray = cv2.equalizeHist(gray)
+    faces = FACE_CASCADE.detectMultiScale(gray, scaleFactor=1.08, minNeighbors=5, minSize=(24, 24))
+    if len(faces) == 0:
+        return None
+    x, y, w, h = max((map(int, face) for face in faces), key=lambda rect: rect[2] * rect[3])
+    return BBox(x=x, y=y, w=w, h=h)
 
 
 def detect_person(image: np.ndarray) -> BBox | None:
@@ -63,10 +106,11 @@ def detect_person(image: np.ndarray) -> BBox | None:
 
 def detect_salient_object(image: np.ndarray) -> BBox | None:
     gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
-    blurred = cv2.GaussianBlur(gray, (7, 7), 0)
-    edges = cv2.Canny(blurred, 40, 120)
-    kernel = np.ones((5, 5), np.uint8)
-    closed = cv2.morphologyEx(edges, cv2.MORPH_CLOSE, kernel, iterations=2)
+    blurred = cv2.GaussianBlur(gray, (9, 9), 0)
+    edges = cv2.Canny(blurred, 30, 110)
+    kernel = np.ones((13, 13), np.uint8)
+    expanded = cv2.dilate(edges, kernel, iterations=1)
+    closed = cv2.morphologyEx(expanded, cv2.MORPH_CLOSE, kernel, iterations=1)
     contours, _ = cv2.findContours(closed, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
     if not contours:
         return None
@@ -77,7 +121,7 @@ def detect_salient_object(image: np.ndarray) -> BBox | None:
     for contour in contours:
         x, y, bw, bh = cv2.boundingRect(contour)
         area = bw * bh
-        if area < max(500, int(image_area * 0.01)):
+        if area < max(1000, int(image_area * 0.015)):
             continue
         candidates.append((area, BBox(x=x, y=y, w=bw, h=bh)))
 
@@ -89,7 +133,7 @@ def detect_salient_object(image: np.ndarray) -> BBox | None:
 
 def square_bbox(bbox: BBox, image_shape: tuple[int, int, int], buffer_ratio: float = 0.15) -> BBox:
     image_h, image_w = image_shape[:2]
-    buffer_ratio = max(0.0, min(buffer_ratio, 0.5))
+    buffer_ratio = max(0.0, min(buffer_ratio, 0.6))
     side = int(round(max(bbox.w, bbox.h) * (1 + buffer_ratio * 2)))
     side = max(1, min(side, image_w, image_h))
 
@@ -127,9 +171,9 @@ def _data_url(image_bytes: bytes, mime_type: str) -> str:
     return f"data:{mime_type};base64,{base64.b64encode(image_bytes).decode('ascii')}"
 
 
-def process_image(image_bytes: bytes, filename: str, buffer_ratio: float = 0.15) -> dict[str, Any]:
+def process_image(image_bytes: bytes, filename: str, buffer_ratio: float = 0.15, detector: str = "auto") -> dict[str, Any]:
     image = decode_image(image_bytes)
-    bbox, method = select_primary_bbox(image)
+    bbox, method = select_primary_bbox(image, detector=detector)
     square = square_bbox(bbox, image.shape, buffer_ratio=buffer_ratio)
     crop = crop_image(image, square)
     annotated = draw_square(image, square)
@@ -139,6 +183,7 @@ def process_image(image_bytes: bytes, filename: str, buffer_ratio: float = 0.15)
 
     return {
         "filename": filename,
+        "detector": detector,
         "method": method,
         "buffer_ratio": buffer_ratio,
         "detected_bbox": bbox.__dict__,