Added face image quality assessment to opencv spaces

README.md

@@ -7,6 +7,13 @@ sdk: gradio
 sdk_version: 5.34.2
 app_file: app.py
 pinned: false
+short_description: Face image quality assessment with ediffiqa using OpenCV
+tags:
+  - opencv
+  - face-image-quality
+  - image-quality-assessment
+  - ediffiqa
+  - yunet
 ---
 
 Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

app.py

@@ -0,0 +1,92 @@
+import cv2 as cv
+import numpy as np
+import gradio as gr
+from huggingface_hub import hf_hub_download
+from yunet import YuNet
+from ediffiqa import eDifFIQA
+
+# Download face detection model (YuNet)
+model_path_yunet = hf_hub_download(
+    repo_id="opencv/face_detection_yunet",
+    filename="face_detection_yunet_2023mar.onnx"
+)
+
+# Download face quality assessment model (eDifFIQA Tiny)
+model_path_quality = hf_hub_download(
+    repo_id="opencv/face_image_quality_assessment_ediffiqa",
+    filename="ediffiqa_tiny_jun2024.onnx"
+)
+
+# Backend and target
+backend_id = cv.dnn.DNN_BACKEND_OPENCV
+target_id = cv.dnn.DNN_TARGET_CPU
+
+# Initialize YuNet for face detection
+face_detector = YuNet(
+    modelPath=model_path_yunet,
+    inputSize=[320, 320],
+    confThreshold=0.9,
+    nmsThreshold=0.3,
+    topK=5000,
+    backendId=backend_id,
+    targetId=target_id
+)
+
+# Initialize eDifFIQA for quality assessment
+quality_model = eDifFIQA(
+    modelPath=model_path_quality,
+    inputSize=[112, 112]
+)
+quality_model.setBackendAndTarget(
+    backendId=backend_id,
+    targetId=target_id
+)
+
+REFERENCE_FACIAL_POINTS = np.array([
+    [38.2946  , 51.6963  ],
+    [73.5318  , 51.5014  ],
+    [56.0252  , 71.7366  ],
+    [41.5493  , 92.3655  ],
+    [70.729904, 92.2041  ]
+], dtype=np.float32)
+
+def align_image(image, detection_data):
+    src_pts = np.float32(detection_data[0][4:-1]).reshape(5, 2)
+    tfm, _ = cv.estimateAffinePartial2D(src_pts, REFERENCE_FACIAL_POINTS, method=cv.LMEDS)
+    face_img = cv.warpAffine(image, tfm, (112, 112))
+    return face_img
+
+def assess_face_quality(input_image):
+    bgr_image = cv.cvtColor(input_image, cv.COLOR_RGB2BGR)
+    h, w, _ = bgr_image.shape
+
+    face_detector.setInputSize([w, h])
+    detections = face_detector.infer(bgr_image)
+
+    if detections is None or len(detections) == 0:
+        return "No face detected.", input_image
+
+    aligned_face = align_image(bgr_image, detections)
+    score = np.squeeze(quality_model.infer(aligned_face)).item()
+
+    output_image = aligned_face.copy()
+    cv.putText(output_image, f"{score:.3f}", (0, 20), cv.FONT_HERSHEY_DUPLEX, 0.8, (0, 0, 255), 2)
+    output_image = cv.cvtColor(output_image, cv.COLOR_BGR2RGB)
+
+    return f"Quality Score: {score:.3f}", output_image
+
+# Gradio Interface
+demo = gr.Interface(
+    fn=assess_face_quality,
+    inputs=gr.Image(type="numpy", label="Upload Face Image"),
+    outputs=[
+        gr.Text(label="Quality Score"),
+        gr.Image(type="numpy", label="Aligned Face with Score")
+    ],
+    title="Face Image Quality Assessment (eDifFIQA + YuNet)",
+    allow_flagging="never",
+    description="Upload a face image. The app detects and aligns the face, then evaluates image quality using the eDifFIQA model."
+)
+
+if __name__ == "__main__":
+    demo.launch()

ediffiqa.py

@@ -0,0 +1,45 @@
+# This file is part of OpenCV Zoo project.
+# It is subject to the license terms in the LICENSE file found in the same directory.
+
+import numpy as np
+import cv2 as cv
+
+
+class eDifFIQA:
+
+    def __init__(self, modelPath, inputSize=[112, 112]):
+        self.modelPath = modelPath
+        self.inputSize = tuple(inputSize) # [w, h]
+
+        self.model = cv.dnn.readNetFromONNX(self.modelPath)
+
+    @property
+    def name(self):
+        return self.__class__.__name__
+
+    def setBackendAndTarget(self, backendId, targetId):
+        self._backendId = backendId
+        self._targetId = targetId
+        self.model.setPreferableBackend(self._backendId)
+        self.model.setPreferableTarget(self._targetId)
+
+    def infer(self, image):
+        # Preprocess image
+        image = self._preprocess(image)
+        # Forward
+        self.model.setInput(image)
+        quality_score = self.model.forward()
+        
+        return quality_score
+
+    def _preprocess(self, image: cv.Mat):
+        # Change image from BGR to RGB
+        image = cv.cvtColor(image, cv.COLOR_BGR2RGB)
+        # Resize to (112, 112)
+        image = cv.resize(image, self.inputSize)
+        # Scale to [0, 1] and normalize by mean=0.5, std=0.5
+        image = ((image / 255) - 0.5) / 0.5
+        # Move channel axis
+        image = np.moveaxis(image[None, ...], -1, 1)
+
+        return image

requirements.txt

@@ -0,0 +1,4 @@
+opencv-python
+gradio
+numpy
+huggingface_hub

yunet.py

@@ -0,0 +1,55 @@
+# This file is part of OpenCV Zoo project.
+# It is subject to the license terms in the LICENSE file found in the same directory.
+#
+# Copyright (C) 2021, Shenzhen Institute of Artificial Intelligence and Robotics for Society, all rights reserved.
+# Third party copyrights are property of their respective owners.
+
+from itertools import product
+
+import numpy as np
+import cv2 as cv
+
+class YuNet:
+    def __init__(self, modelPath, inputSize=[320, 320], confThreshold=0.6, nmsThreshold=0.3, topK=5000, backendId=0, targetId=0):
+        self._modelPath = modelPath
+        self._inputSize = tuple(inputSize) # [w, h]
+        self._confThreshold = confThreshold
+        self._nmsThreshold = nmsThreshold
+        self._topK = topK
+        self._backendId = backendId
+        self._targetId = targetId
+
+        self._model = cv.FaceDetectorYN.create(
+            model=self._modelPath,
+            config="",
+            input_size=self._inputSize,
+            score_threshold=self._confThreshold,
+            nms_threshold=self._nmsThreshold,
+            top_k=self._topK,
+            backend_id=self._backendId,
+            target_id=self._targetId)
+
+    @property
+    def name(self):
+        return self.__class__.__name__
+
+    def setBackendAndTarget(self, backendId, targetId):
+        self._backendId = backendId
+        self._targetId = targetId
+        self._model = cv.FaceDetectorYN.create(
+            model=self._modelPath,
+            config="",
+            input_size=self._inputSize,
+            score_threshold=self._confThreshold,
+            nms_threshold=self._nmsThreshold,
+            top_k=self._topK,
+            backend_id=self._backendId,
+            target_id=self._targetId)
+
+    def setInputSize(self, input_size):
+        self._model.setInputSize(tuple(input_size))
+
+    def infer(self, image):
+        # Forward
+        faces = self._model.detect(image)
+        return np.empty(shape=(0, 5)) if faces[1] is None else faces[1]