1 files changed, 77 insertions, 0 deletions
diff --git a/megapixels/notebooks/utils/HaarSaliency.py b/megapixels/notebooks/utils/HaarSaliency.py
new file mode 100644
index 00000000..6a511efb
--- /dev/null
+++ b/megapixels/notebooks/utils/HaarSaliency.py
@@ -0,0 +1,77 @@
+import os
+from os.path import join
+import cv2
+import numpy as np
+import xml.etree.ElementTree
+class HaarSaliency:
+
+    # Input images should be between 320x240 and 640x480
+    classifiers = []
+    face_matrix = []
+    flags = 0
+    nstages = 0
+
+    def __init__(self, cascade, min_size=60, max_size=400,
+        face_neighbors=3, sal_neighbors=0, blur_kernel=(31,31), scale_factor=1.1, stage_start=1):
+
+        self.face_neighbors = face_neighbors
+        self.sal_neighbors = sal_neighbors
+        self.scale_factor = scale_factor
+        self.blur_kernel = blur_kernel
+        self.min_size = (min_size, min_size)
+        self.max_size = (max_size, max_size)
+        self.stage_start = stage_start
+
+        cdir, cname = os.path.split(cascade)
+        cname, ext = os.path.splitext(cname)
+
+        root = xml.etree.ElementTree.parse(cascade)
+        #width = int(root.find('./cascade/width').text.strip())
+        #height = int(root.find('./cascade/height').text.strip())
+        self.nstages = int(root.find('./cascade/stageNum').text.strip())
+
+        # initialize classifiers
+        cascades = [join(cdir,cname,str(c)+'.xml') for c in range(stage_start,self.nstages+1) ]
+        self.classifiers = [cv2.CascadeClassifier(c) for c in cascades]
+
+    def get_saliency(self,src):
+        
+        # conver to grayscale
+        src_gray = cv2.cvtColor(src,cv2.COLOR_BGR2GRAY)
+
+        # run face detector on all stage-classifiers
+        self.face_matrix = [self.detect_faces(src_gray,c) for c in self.classifiers]
+
+        # create saliency map in grayscale
+        w,h = src.shape[1],src.shape[0]
+        saliency = np.zeros((h,w), dtype=np.float32).reshape(h,w)
+
+        # draw face-roi as stage-weighted grayscale vals
+        # min_neighbors sets max value for grayscale --> white
+        for i,face_list in enumerate(self.face_matrix,1):
+            inc = round(255./float(self.face_neighbors)/float(self.nstages-self.stage_start)) * i
+            if face_list is not None:
+                for x1,y1,fw,fh in face_list:
+                    saliency[y1:y1+fh,x1:x1+fw] += inc
+
+        # normalize, clip, and recast as uint8
+        smax = saliency.max()
+        if smax > 255:
+            saliency /= (smax/255)
+        saliency = np.clip(saliency,0,255)
+        saliency = np.array(saliency,dtype=np.uint8)
+        
+        # blur, colormap, and composite
+        saliency = cv2.GaussianBlur(saliency,self.blur_kernel,0)
+        dst = cv2.applyColorMap(saliency, cv2.COLORMAP_JET)
+        return dst
+
+    def detect_faces(self,src,classifier):
+
+        matches = classifier.detectMultiScale(src, self.scale_factor,
+            self.sal_neighbors, self.flags, self.min_size, self.max_size)
+
+        if matches is None or len(matches) < 1:
+            return None
+        else:
+            return sorted(map(tuple,matches),reverse=True) # lg --> sm
+\ No newline at end of file