Initial Commit

src/LBP.py

@@ -0,0 +1,35 @@
+import cv2
+from skimage import feature
+import matplotlib.pyplot as plt
+
+class LBPImageEncoder:
+    def __init__(self, numPoints, radius):
+        self.numPoints = numPoints
+        self.radius = radius
+
+    def describe(self, image, eps=1e-7):
+        lbp = feature.local_binary_pattern(image, self.numPoints, self.radius)
+        hist = plt.hist(lbp.ravel())
+        return lbp, hist
+
+    def face_detection(image):
+        cascadePath = "haarcascade_frontalface_default.xml"
+        detector = cv2.CascadeClassifier(cascadePath)
+
+        gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
+        rects = detector.detectMultiScale(gray, scaleFactor=1.05, minNeighbors=10, minSize=(30, 30),flags=cv2.CASCADE_SCALE_IMAGE)
+        return rects
+
+    def preprocess_img(self, imagePath):
+        img = cv2.imread(imagePath)
+        rects = self.face_detection(img)
+        for (x, y, w, h) in rects:
+            face = img[y:y + h, x:x + w]
+            face = cv2.cvtColor(face, cv2.COLOR_BGR2GRAY)
+        # plt.imshow(face , cmap="gray")
+        # print(face.shape)
+        # face = np.array(face)
+
+        lbp, hist = self.describe(face)
+        return lbp, hist
+

src/app.py

@@ -1,5 +1,6 @@
 from colordescriptor import ColorDescriptor
 from CLIP import CLIPImageEncoder
+from LBP import LBPImageEncoder
 import gradio as gr
 import os
 import cv2
@@ -22,11 +23,15 @@ def index_dataset(dataset):
     clip_model = CLIPImageEncoder()
     dataset_with_embeddings = dataset_with_embeddings.map(clip_model.encode_images, batched=True, batch_size=16)
 
+    ## LBP Embeddings
+    lbp_model = LBPImageEncoder()
+    dataset_with_embeddings = dataset_with_embeddings.map(lambda row: {'lbp_embeddings': lbp_model.describe(row["image"])})
+
     # Add index
     dataset_with_embeddings.add_faiss_index(column='color_embeddings')
     dataset_with_embeddings.save_faiss_index('color_embeddings', 'color_index.faiss')
-
     dataset_with_embeddings.add_faiss_index(column='clip_embeddings')
+    dataset_with_embeddings.add_faiss_index(column='lbp_embeddings')
     dataset_with_embeddings.save_faiss_index('clip_embeddings', 'clip_index.faiss')
 
 
@@ -47,6 +52,7 @@ def check_index(ds):
 dataset_with_embeddings = check_index(candidate_subset)
 
 # Main function, to find similar images
+# TODO: allow different descriptor/embedding functions
 # TODO: implement different distance measures
 
 def get_neighbors(query_image, selected_descriptor, top_k=5):
@@ -74,6 +80,13 @@ def get_neighbors(query_image, selected_descriptor, top_k=5):
             'clip_embeddings', qi_embedding, k=top_k)
         images = retrieved_examples['image']
         return images
+    if "LBP" in selected_descriptor:
+        lbp_model = LBPImageEncoder(8,2)
+        qi_embedding = lbp_model.preprocess_img(query_image)
+        scores, retrieved_examples = dataset_with_embeddings.get_nearest_examples(
+            'lbp_embeddings', qi_embedding, k=top_k)
+        images = retrieved_examples['image']
+        return images
     else:
         print("This descriptor is not yet supported :(")
         return []