add the predict threshold, change the directory architecture

jiangyiqiao · Mar 30, 2018 · fe9027a · fe9027a
1 parent fdaee7d
commit fe9027a
Show file tree

Hide file tree

Showing 33 changed files with 92 additions and 42 deletions.
diff --git a/README.md b/README.md
@@ -1,7 +1,7 @@
-# Politicians_FaceRecongnise
+# Politicians_FaceRecognise
 
 ## Introduction
-This is a project to realise 55 Chinese politicians face recongnise. refernce the repository ([facenet](https://github.com/davidsandberg/facenet.git)).
+This is a project to realise **55 Chinese politicians** face recognise. refernce the repository **([facenet](https://github.com/davidsandberg/facenet.git))**.
 ## Pre-trained models
 | Model name      | LFW accuracy | Training dataset | Architecture |
 |-----------------|--------------|------------------|-------------|
@@ -13,41 +13,65 @@ This is a project to realise 55 Chinese politicians face recongnise. refernce th
 The code is tested using Tensorflow 0.12 under Ubuntu 16.04 with Python3.6 and Python3.5
 * tensorflow>0.12
 * sklearn
-* matpl
 * numpy
 * scipy
 * pickle
 * cv2
 * matplotlib
 
+I set the code:
+
+    CLASS_PROBABILITY_THRESHOLD=0.2    #the probability of the predict threshold，if less than the threshold ,set the prediction="unknown"
+
+## Dataset
+1. for train set: 
+
+    images/train/aligned_policy/   #person num: 55 + others, pictures num: 2409
+
+2. for test dataset:
+
+    images/test/others or images/test/policy
+
+3. for train the model:
+
+    python train_model.py
+
+4. for test the model:
+
+    python test_model  # I didn't set the predict threshod,the result will output the max probobility of classname. 
+
+
 if you want to change the pictures ,in order to use your own data,also if someone is interested to use my dataset about 55 Chinese politicians,you can email to me ,I am glad to share you my dataset. 
 1. put your images that haven't be aligned into the directory align/images/,like:
-   ```
+
    align/images/policy/
          people1/
                1.jpg
                2.jpg
          people2/
                1.jpg
                2.jpg
-   ```
+
 
 2. you can change the input or output directory 
-    ```
+
     parser.add_argument('--input_dir', type=str, help='Directory with unaligned images.',default='images/policy/')
     parser.add_argument('--output_dir', type=str, help='Directory with aligned face thumbnails.',default='images/aligned_policy/')
-    ```
+
 
 then run the code
 
-    ```
+
     python align_dataset_mtcnn.py
-    ```
+
+
+after run this code ,you will get the anigned_pictures,you can change the parameters to choose if you want to detect_multiple_faces,set:
 
-after run this code ,you will get the anigned_pictures,you can change the parameters to choose if you want to detect_multiple_faces,the result like:
+    parser.add_argument('--detect_multiple_faces', type=bool, help='Detect and align multiple faces per image.', default=True)
 
-    ```
-    align/images/aligned_policy/
+the result like:
+
+    align/images/train/aligned_policy/
         people1/
                1.jpg
                2.jpg
@@ -56,37 +80,37 @@ after run this code ,you will get the anigned_pictures,you can change the parame
                1.jpg
                1_1.jpg
                2.jpg
-    ```
+
 
-3. copy the files align/images/aligned_policy into images/ 
+3. copy the files align/images/aligned_policy into images/train/ 
 
 if you want to use my model directly,and run my project and see the result, you can
 
 1. show the politician pictures and see the prediction or show the other people which is not the politicican one by one:
-    ```
+
     python calacc_plt.py
     python calerror_plt.py
-    ```
+
 
 2. I alse provide the multi thread python code to calculate the accuracy.
-    ```
+
     python multiThread_process.py
-    ```
+
 
 # Results
-1. can recongnise profile 
+1. can recognise profile 
 
 ![Figure_1](/result/Figure_1.png)
 
 2. alse has some error
 
 ![Figure_1-1](/result/Figure_1-1.png)
 
-3. can recongise sepcial part face
+3. can recogise sepcial part face
 
 ![Figure_1-2](/result/Figure_1-2.png)
 
-4. can recongise sepcial part face
+4. can recogise sepcial part face
 
 ![Figure_1-3](/result/Figure_1-3.png)
 

diff --git a/__pycache__/facenet.cpython-35.pyc b/__pycache__/facenet.cpython-35.pyc
diff --git a/align/__pycache__/__init__.cpython-35.pyc b/align/__pycache__/__init__.cpython-35.pyc
diff --git a/align/__pycache__/detect_face.cpython-35.pyc b/align/__pycache__/detect_face.cpython-35.pyc
diff --git a/calacc_plt.py b/calacc_plt.py
@@ -21,6 +21,7 @@
 from sklearn.svm import SVC
 
 count_rightpic=0
+CLASS_PROBABILITY_THRESHOLD=0.2           #设置阈值准确度，否则识别为unknown
 def main(args):
     dataset = facenet.get_dataset(args.input_dir)
 
@@ -56,14 +57,21 @@ def main(args):
                         feed_dict = {images_placeholder: [np.array(image)], phase_train_placeholder: False}
                         emb_datas = sess.run(embeddings, feed_dict=feed_dict)
 
-                        print("emb_datas")
-                        #print(emb_datas)
                         print('Testing classifier')
-                        prediction = model.predict(emb_datas)
-                        print(prediction)
-                        predictions.append(class_names[prediction[0]])
-
-            #每张图所有预测人脸结果
+                        predict_proba=model.predict_proba(emb_datas)
+                        best_class_indice = np.argmax(predict_proba, axis=1)
+                        best_class_probability = predict_proba[np.arange(len(best_class_indice)), best_class_indice]
+                        for i in range(len(best_class_indice)):
+                            class_name = class_names[best_class_indice[i]]
+                            class_probability = best_class_probability[i]
+                            print("class_probability:",class_probability)
+                            if class_probability < CLASS_PROBABILITY_THRESHOLD:
+                               prediction = "unknown"
+                            else:
+                               prediction=class_name
+                               break
+                        predictions.append(prediction)
+
             for prediction in predictions:
                 print("prediction:%s"%prediction)
 

diff --git a/calerror_plt.py b/calerror_plt.py
@@ -28,6 +28,7 @@
 'jiangzeming','lipeng','maozedong','wenjiabao','zhaoziyang','zhurongji','boxilai','guoboxiong','linjihua',
 'xucaihou','zhouyongkang']
 
+CLASS_PROBABILITY_THRESHOLD=0.2           #设置阈值准确度，否则识别为unknown
 
 def main(args):
     dataset = facenet.get_dataset(args.input_dir)
@@ -63,13 +64,21 @@ def main(args):
                     for image in images:
                         feed_dict = {images_placeholder: [np.array(image)], phase_train_placeholder: False}
                         emb_datas = sess.run(embeddings, feed_dict=feed_dict)
-
-                        print("emb_datas")
-                        #print(emb_datas)
+
                         print('Testing classifier')
-                        prediction = model.predict(emb_datas)
-                        print(prediction)
-                        predictions.append(class_names[prediction[0]])
+                        predict_proba=model.predict_proba(emb_datas)
+                        best_class_indice = np.argmax(predict_proba, axis=1)
+                        best_class_probability = predict_proba[np.arange(len(best_class_indice)), best_class_indice]
+                        for i in range(len(best_class_indice)):
+                            class_name = class_names[best_class_indice[i]]
+                            class_probability = best_class_probability[i]
+                            print("class_probability:",class_probability)
+                            if class_probability < CLASS_PROBABILITY_THRESHOLD:
+                               prediction = "unknown"
+                            else:
+                               prediction=class_name
+                               break
+                        predictions.append(prediction)
 
             #每张图所有预测人脸结果
             for prediction in predictions:

diff --git a/...thers/old/0-31p2021702a339957500-s_73.jpg → ...thers/old/0-31p2021702a339957500-s_73.jpg b/...thers/old/0-31p2021702a339957500-s_73.jpg → ...thers/old/0-31p2021702a339957500-s_73.jpg
diff --git a/images/others/old/0.jpg → images/test/others/old/0.jpg b/images/others/old/0.jpg → images/test/others/old/0.jpg
diff --git a/images/others/old/1.jpg → images/test/others/old/1.jpg b/images/others/old/1.jpg → images/test/others/old/1.jpg
diff --git a/...hers/old/1f24de4448ac969300a65de7c_12.jpg → ...hers/old/1f24de4448ac969300a65de7c_12.jpg b/...hers/old/1f24de4448ac969300a65de7c_12.jpg → ...hers/old/1f24de4448ac969300a65de7c_12.jpg
diff --git a/...rs/young/0-20171223151400370816313_31.jpg → ...rs/young/0-20171223151400370816313_31.jpg b/...rs/young/0-20171223151400370816313_31.jpg → ...rs/young/0-20171223151400370816313_31.jpg
diff --git a/...rs/young/0-20171223151400380481651_29.jpg → ...rs/young/0-20171223151400380481651_29.jpg b/...rs/young/0-20171223151400380481651_29.jpg → ...rs/young/0-20171223151400380481651_29.jpg
diff --git a/...ers/young/0-20171223151400390719803_4.jpg → ...ers/young/0-20171223151400390719803_4.jpg b/...ers/young/0-20171223151400390719803_4.jpg → ...ers/young/0-20171223151400390719803_4.jpg
diff --git a/images/policy/xijinping/1.jpeg → images/test/policy/xijinping/1.jpeg b/images/policy/xijinping/1.jpeg → images/test/policy/xijinping/1.jpeg
diff --git a/images/policy/xijinping/2.jpeg → images/test/policy/xijinping/2.jpeg b/images/policy/xijinping/2.jpeg → images/test/policy/xijinping/2.jpeg
diff --git a/images/policy/xijinping/3.jpeg → images/test/policy/xijinping/3.jpeg b/images/policy/xijinping/3.jpeg → images/test/policy/xijinping/3.jpeg
diff --git a/images/policy/xijinping/4.jpeg → images/test/policy/xijinping/4.jpeg b/images/policy/xijinping/4.jpeg → images/test/policy/xijinping/4.jpeg
diff --git a/images/policy/xijinping/5.jpeg → images/test/policy/xijinping/5.jpeg b/images/policy/xijinping/5.jpeg → images/test/policy/xijinping/5.jpeg
diff --git a/images/policy/xijinping/6.jpeg → images/test/policy/xijinping/6.jpeg b/images/policy/xijinping/6.jpeg → images/test/policy/xijinping/6.jpeg
diff --git a/images/policy/xijinping/7.jpeg → images/test/policy/xijinping/7.jpeg b/images/policy/xijinping/7.jpeg → images/test/policy/xijinping/7.jpeg
diff --git a/images/policy/xijinping/8.jpeg → images/test/policy/xijinping/8.jpeg b/images/policy/xijinping/8.jpeg → images/test/policy/xijinping/8.jpeg
diff --git a/images/policy/xijinping/9.jpeg → images/test/policy/xijinping/9.jpeg b/images/policy/xijinping/9.jpeg → images/test/policy/xijinping/9.jpeg
diff --git a/images/policy/xijinping/下载.jpeg → images/test/policy/xijinping/下载.jpeg b/images/policy/xijinping/下载.jpeg → images/test/policy/xijinping/下载.jpeg
diff --git a/...jinping/003f44d305ea4db1b3d5cdd1_32_0.png → ...jinping/003f44d305ea4db1b3d5cdd1_32_0.png b/...jinping/003f44d305ea4db1b3d5cdd1_32_0.png → ...jinping/003f44d305ea4db1b3d5cdd1_32_0.png
diff --git a/images/aligned_policy/xijinping/14_1.png → ...s/train/aligned_policy/xijinping/14_1.png b/images/aligned_policy/xijinping/14_1.png → ...s/train/aligned_policy/xijinping/14_1.png
diff --git a/images/aligned_policy/xijinping/15_3.png → ...s/train/aligned_policy/xijinping/15_3.png b/images/aligned_policy/xijinping/15_3.png → ...s/train/aligned_policy/xijinping/15_3.png
diff --git a/...inping/16japanchina-2-master105_134_7.png → ...inping/16japanchina-2-master105_134_7.png b/...inping/16japanchina-2-master105_134_7.png → ...inping/16japanchina-2-master105_134_7.png
diff --git a/...inping/3E-9945A72E672D_w1023_r1__31_0.png → ...inping/3E-9945A72E672D_w1023_r1__31_0.png b/...inping/3E-9945A72E672D_w1023_r1__31_0.png → ...inping/3E-9945A72E672D_w1023_r1__31_0.png
diff --git a/images/aligned_policy/xijinping/3_0.png → ...es/train/aligned_policy/xijinping/3_0.png b/images/aligned_policy/xijinping/3_0.png → ...es/train/aligned_policy/xijinping/3_0.png
diff --git a/...nping/4succession-1-articleLarg_167_1.png → ...nping/4succession-1-articleLarg_167_1.png b/...nping/4succession-1-articleLarg_167_1.png → ...nping/4succession-1-articleLarg_167_1.png
diff --git a/multiThread_process.py b/multiThread_process.py
@@ -35,6 +35,7 @@
 batch_size=90
 facenet_image_size=160
 
+CLASS_PROBABILITY_THRESHOLD=0.2           #设置阈值准确度，否则识别为unknown
 
 def parsePicture(picture_paths,model,class_names):
     num=0
@@ -78,12 +79,20 @@ def embeding(image_path,model,class_names):
             for image in images:
                 feed_dict = {images_placeholder: [np.array(image)], phase_train_placeholder: False}
                 emb_datas = sess.run(embeddings, feed_dict=feed_dict)
-                print("emb_datas")
-
                 print('Testing classifier')
-                prediction = model.predict(emb_datas)
-                print(prediction)
-                predictions.append(class_names[prediction[0]])
+                predict_proba=model.predict_proba(emb_datas)
+                best_class_indice = np.argmax(predict_proba, axis=1)
+                best_class_probability = predict_proba[np.arange(len(best_class_indice)), best_class_indice]
+                for i in range(len(best_class_indice)):
+                    class_name = class_names[best_class_indice[i]]
+                    class_probability = best_class_probability[i]
+                    print("class_probability:",class_probability)
+                    if class_probability < CLASS_PROBABILITY_THRESHOLD:
+                        prediction = "unknown"
+                    else:
+                        prediction=class_name
+                        break
+                predictions.append(prediction)
 
         #每张图所有预测人脸结果
     for prediction in predictions:

diff --git a/test_model.py b/test_model.py
@@ -60,7 +60,7 @@ def main(args):
                         print('Testing classifier')
                         prediction = model.predict(emb_datas)    #预测结果
                         print(prediction)
-                        predictions.append(class_names[prediction[0]])  
+                        predictions.append(class_names[prediction[0]])    #直接预测，总会得到最大的可能值
 
 
             for prediction in predictions:         #每张图所有预测人脸结果
@@ -134,7 +134,7 @@ def load_and_align_data(image_path, image_size, margin, gpu_memory_fraction):
 def parse_arguments(argv):
     parser = argparse.ArgumentParser()
 
-    parser.add_argument('--input_dir', type=str, help='Directory with unaligned images.',default='./images/test/')
+    parser.add_argument('--input_dir', type=str, help='Directory with unaligned images.',default='./images/test/policy')
     parser.add_argument('--image_size', type=int,
         help='Image size (height, width) in pixels.', default=182)
     parser.add_argument('--margin', type=int,

diff --git a/train_model.py b/train_model.py
@@ -101,7 +101,7 @@ def parse_arguments(argv):
     parser = argparse.ArgumentParser()
 
     parser.add_argument('--data_dir', type=str,
-        help='Path to the data directory containing aligned LFW face patches.',default='images/aligned_policy/')
+        help='Path to the data directory containing aligned LFW face patches.',default='images/train/aligned_policy/')
     parser.add_argument('--model', type=str, 
         help='Could be either a directory containing the meta_file and ckpt_file or a model protobuf (.pb) file',default='models/policy/embeding.pb')
     parser.add_argument('--classifier_filename',