DeepFace -> face_recognition (dlib) migration (#70)

* feat,ref: swapped register faces to register embeddings

* feat: added func for iterating over embeddings

* feat: added register face embeddings func and face match for dlib face-rec

* fix: robustness fixes for faces folder and get_facematch

* build: added face-recognition to poetry config

* ref,doc: Cleaned up recognition after dlib changes

* ref,doc: added documentation for iter_face_embeddings

* fix: put rpi-gpio back

* feat: face_unlock.py migrated and dlib working on laptop

* feat: Made face_unlock.py work for laptop or rpi

* feat: made statuses more specific and added Enum

* fix: restored take_snapshot

Opencv only takes next frame in buffer not most recent so we need to reset device each snapshot

* ref: deleted old test scripts as these are captured by face_unlock.py now

* ref: removed explore-deepface script

client/data/routines.py

@@ -7,10 +7,9 @@
 import sqlite3
 
 from datetime import datetime
-from typing import Any, NamedTuple, Optional
+from typing import Any, NamedTuple, Optional, Iterator
 from importlib import resources
 
-import cv2
 import numpy as np
 from pydbml import PyDBML
 
@@ -186,20 +185,30 @@ def get_user_postures(
         return [Posture(*record) for record in result.fetchall()]
 
 
-def register_faces(user_id: int, faces: list[np.ndarray]) -> None:
-    """Register faces for a user.
+def register_face_embeddings(user_id: int, face_embeddings: list[np.ndarray]) -> None:
+    """Register face embeddings for a user.
 
     Args:
         user_id: The user to register faces for.
-        faces: List of face arrays in the format HxWxC where channels are RGB
+        faces: List of face embedding arrays.
     """
+    stacked_faces = np.vstack(face_embeddings)
     with resources.as_file(FACES_FOLDER) as faces_folder:
-        faces_folder.mkdir(exist_ok=True)
-        user_folder = faces_folder / str(user_id)
-        user_folder.mkdir()
-        for i, image in enumerate(faces):
-            image_path = user_folder / f"{i}.png"
-            cv2.imwrite(str(image_path), image)
+        embedding_path = faces_folder / f"{user_id}.npy"
+        np.save(embedding_path, stacked_faces)
+
+
+def iter_face_embeddings() -> Iterator[tuple[int, list[np.ndarray]]]:
+    """
+    Returns:
+        Generator which yields (user_id, face_embeddings) each iteration. Each iteration will give
+            a different user. face_embeddings is a list of numpy arrays which each represent an
+            embedded face for the user.
+    """
+    with resources.as_file(FACES_FOLDER) as faces_folder:
+        for user_embeddings_path in faces_folder.iterdir():
+            user_id = int(user_embeddings_path.stem)
+            yield user_id, list(np.load(user_embeddings_path))
 
 
 def get_schema_info() -> list[list[tuple[Any]]]:
@@ -228,3 +237,11 @@ def _connect() -> sqlite3.Connection:
         return sqlite3.connect(
             database_file, detect_types=sqlite3.PARSE_DECLTYPES | sqlite3.PARSE_COLNAMES
         )
+
+
+def _init_faces_folder() -> None:
+    with resources.as_file(FACES_FOLDER) as faces_folder:
+        faces_folder.mkdir(exist_ok=True)
+
+
+_init_faces_folder()

client/models/face_recognition/recognition.py

@@ -1,14 +1,18 @@
-from importlib import resources
+from enum import Enum
 
 import numpy as np
-from data.routines import FACES_FOLDER
-from deepface import DeepFace
+import face_recognition
 
-MODEL_NAME = "GhostFaceNet"
+from data.routines import register_face_embeddings, iter_face_embeddings
 
+MODEL_NAME = "small"
 
-def _path_to_user_id(path: str) -> str:
-    return path.split("/")[-2]
+
+class Status(Enum):
+    NO_FACES = -3
+    TOO_MANY_FACES = -2
+    NO_MATCH = -1
+    OK = 0
 
 
 def get_face_match(login_face: np.ndarray) -> int:
@@ -19,13 +23,50 @@ def get_face_match(login_face: np.ndarray) -> int:
         login_face: Image of user's face as an array
 
     Returns:
-        Matching user id, or -1 if no users matched
+        Matching user id, or one of Status values.
     """
-    with resources.as_file(FACES_FOLDER) as faces_folder:
-        try:
-            dfs = DeepFace.find(login_face, str(faces_folder), model_name=MODEL_NAME)
-        except ValueError:
-            return -1
-        df = dfs[0]
-        user_id = int(_path_to_user_id(df.iloc[0]["identity"]))
-        return user_id
+    login_embeddings = face_recognition.face_encodings(login_face, model=MODEL_NAME)
+
+    # Should only detect exactly one face
+    if len(login_embeddings) == 0:
+        return Status.NO_FACES.value
+    if len(login_embeddings) > 1:
+        return Status.TOO_MANY_FACES.value
+
+    login_embedding = login_embeddings[0]
+
+    for user_id, user_embeddings in iter_face_embeddings():
+        matches = face_recognition.compare_faces(user_embeddings, login_embedding)
+
+        if any(matches):
+            return user_id
+
+    return Status.NO_MATCH.value
+
+
+def register_faces(user_id: int, faces: list[np.ndarray]) -> int:
+    """Compute and store face embeddings in the database.
+
+    Args:
+        user_id: Id of the user who belongs to the faces
+        faces: List of face images in the shape HxWxC where (C)hannels are in RGB
+
+    Returns:
+        Registration status
+    """
+    face_embeddings = []
+    for face in faces:
+        all_faces_embed = face_recognition.face_encodings(face, model=MODEL_NAME)
+
+        # Should only detect exactly one face
+        if len(all_faces_embed) == 0:
+            return Status.NO_FACES.value
+        if len(all_faces_embed) > 1:
+            return Status.TOO_MANY_FACES.value
+
+        face_embedding = all_faces_embed[0]
+        face_embeddings.append(face_embedding)
+
+    register_face_embeddings(user_id, face_embeddings)
+
+    return Status.OK.value