updated app: can now detect and label multiple faces in a given image

app.py

@@ -1,3 +1,4 @@
+import cv2
 import os
 import gradio as gr
 import numpy as np
@@ -8,23 +9,40 @@ from tensorflow.image import resize
 from PIL import Image
 
 #gradio interface
-def classify_image(image):
-    input_arr = img_to_array(image)/255 #convert PIL object to numpy array and normalize
-    input_arr_resh = resize(input_arr, (48, 48)).numpy()
-    if model.channelno == 1:
-        # Model expects inputs of shape (48,48,1)
-        input_arr_resh_gray = input_arr_resh.mean(axis=2).reshape(48,48,1)
-        predictions = model.predictor.predict(np.array([input_arr_resh_gray]))
-    elif model.channelno == 3:
-        # Model expects inputs of shape (48,48,3)
-        input_arr_resh_4dims = np.expand_dims(input_arr_resh, axis=0)
-        predictions = model.predictor.predict(input_arr_resh_4dims)
-    pr_emotion = model.labeldict[predictions.argmax()]
-    prob = predictions.max()*100
-    returnstr = f'Prediction: {pr_emotion}, probability: {prob:4.1f}%'
-    predictions_f = ['%s:%5.2f'%(model.labeldict[i],p*100) for i,p in enumerate(predictions[0])]
-    print(predictions_f)
-    return returnstr
+def process_image(img):
+    cv2_face_cascade = cv2.CascadeClassifier(cv2.data.haarcascades + 
+                                             'haarcascade_frontalface_default.xml')
+    img, _ = annotate_objects_in_image(img, 
+                                       cv2_face_cascade, 
+                                       model.labeldict, 
+                                       model.predictor)
+
+    return(img)
+
+def annotate_objects_in_image(img, obj_cascade, labeldict, model):
+    #img = cv2.flip(img, 1)
+    gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
+    objects = obj_cascade.detectMultiScale(gray, 1.1, 4)
+    model_result = ""
+    try:
+        for (x, y, w, h) in objects:
+            face = img[y-50:y+h+50, x-50:x+w+50]
+            model_result = predict_emotion_from_image(face, labeldict, model)
+            font = cv2.FONT_HERSHEY_SIMPLEX
+            cv2.putText(img, model_result, (x, y), font, 1, (0, 0, 255), 2, cv2.LINE_AA)
+            cv2.rectangle(img, (x-25, y-25), (x+w+25, y+h+25), (0, 255, 0), 2)
+    except Exception as ex:
+        model_result = ""
+    return img, model_result
+    
+def predict_emotion_from_image (face_raw, labeldict, model):
+    face = cv2.resize(face_raw, (48, 48))
+    face_array = np.array(face)/255
+    face_array_4dims = np.expand_dims(face_array, axis=0)
+    prediction_vec = model.predict(face_array_4dims)
+    prediction = np.argmax(prediction_vec)
+    emotion = labeldict[prediction]
+    return (emotion)
 
 class ModelClass:
     def __init__(self,name='EDA_CNN.h5'):
@@ -45,17 +63,12 @@ modeltouse = "MobileNet12blocks_wdgenaug_onrawdata_valacc063.h5"
 
 model = ModelClass(modeltouse)
 
-image = gr.inputs.Image(shape=(48,48))
-label = gr.outputs.Label()
-examples = ['Happy_48_48_%d.png'%model.channelno,
-            'Neutral_48_48_%d.png'%model.channelno,
-            'Fear_48_48_%d.png'%model.channelno,
-            'Angry_48_48_%d.png'%model.channelno,
-            'Sad_48_48_%d.png'%model.channelno, 
-            #'Disgust_48_48_%d.png'%model.channelno, 
-            'Surprise_48_48_%d.png'%model.channelno]
-
-# image = Image.open('./Happy_48_48_%d.png'%model.channelno)
-# classify_image(image)
-intf = gr.Interface(fn=classify_image, inputs=image, outputs=label, examples=examples)
+image_in = gr.inputs.Image() #shape=(48,48)
+image_out = gr.inputs.Image() 
+examples = ['OnurH_CerenH.jpg', 'OnurA_CerenH.jpg']
+#fname = 'Onur_happy.jpg'
+#image = cv2.imread(fname)
+#process_image(image, model)
+
+intf = gr.Interface(fn=process_image, inputs=image_in, outputs=image_out, examples=examples)
 intf.launch(inline=False)