有幸參加了微軟OpenHack挑戰賽,雖然題目難度不大,但是很有意思,學到了很多東西,還有幸認識了微軟樑健老師,謝謝您的幫助!同時還認識同行的很多朋友,非常高興,把這段難忘的比賽記錄一下~~也分享一下程式碼,給那些沒有參加的朋友,
資料集(文末連結)
首先每支隊伍會收到一個資料集,它是一個登山公司提供的裝備圖片,有登山鎬,鞋子,登山扣,不知道叫什麼的雪地爪?手套,衝鋒衣,安全帶。。。一共12個類別,每個類別幾百個樣本,我們的任務就是對這些圖片分類和識別
簡單看一下:
賽題:
賽題共有6道,簡單描述一下:
1、搭建環境(略過)
2、影像正規化(包括顏色和大小)
3、通過機器學習方法對影像分類,precision>0.8
4、通過深度學習方法對影像分類,precision>0.9
5、部署(略過)
6、目標檢測(用全新的資料集,檢測雪地中的登山者是否帶頭盔!!航拍影像,有點難度~)
_______________________________________
下面是每道題目的詳細描述和程式碼
題目2
完成以下任務:
選擇一種基本顏色,例如白色並填充所有圖片尺寸不是1:1比例的影像
不通過直接拉伸的方式,重塑至128x128x3畫素的陣列形狀
確保每個影像的畫素範圍從0到255(包含或[0,255]),也稱為“對比度拉伸”(contrast stretching).
標準化或均衡以確保畫素在[0,255]範圍內.
成功完成的標準
團隊將在Jupyter Notebook中執行一個程式碼單元,繪製原始影像,然後繪製填充後的畫素值歸一化或均衡影像, 展示給教練看.
團隊將在Jupyter notebook 為教練執行一個程式碼單元,顯示的畫素值的直方圖應該在0到255的範圍內(包括0和255).
def normalize(src): arr = array(src) arr = arr.astype('float') # Do not touch the alpha channel for i in range(3): minval = arr[...,i].min() maxval = arr[...,i].max() if minval != maxval: arr[...,i] -= minval arr[...,i] *= (255.0/(maxval-minval)) arr = arr.astype(uint8) return Image.fromarray(arr,'RGB') import matplotlib.pyplot as plt from PIL import ImageColor from matplotlib.pyplot import imshow from PIL import Image from pylab import * import copy plt.figure(figsize=(10,10)) #設定視窗大小 # src = Image.open("100974.jpeg") src = Image.open("rose.jpg") src_array = array(src) plt.subplot(2,2,1), plt.title('src') plt.imshow(src), plt.axis('off') ar=src_array[:,:,0].flatten() ag=src_array[:,:,1].flatten() ab=src_array[:,:,2].flatten() plt.subplot(2,2,2), plt.title('src hist') plt.axis([0,255,0,0.03]) plt.hist(ar, bins=256, normed=1,facecolor='red',edgecolor='r',hold=1) #原始影像直方圖 plt.hist(ag, bins=256, normed=1,facecolor='g',edgecolor='g',hold=1) #原始影像直方圖 plt.hist(ab, bins=256, normed=1,facecolor='b',edgecolor='b') #原g始影像直方圖 dst = normalize(src) dst_array = array(dst) plt.subplot(2,2,3), plt.title('dst') plt.imshow(dst), plt.axis('off') ar=dst_array[:,:,0].flatten() ag=dst_array[:,:,1].flatten() ab=dst_array[:,:,2].flatten() plt.subplot(2,2,4), plt.title('dst hist') plt.axis([0,255,0,0.03]) plt.hist(ar, bins=256, normed=1,facecolor='red',edgecolor='r',hold=1) #原始影像直方圖 plt.hist(ag, bins=256, normed=1,facecolor='g',edgecolor='g',hold=1) #原始影像直方圖 plt.hist(ab, bins=256, normed=1,facecolor='b',edgecolor='b') #原g始影像直方圖
題目3
使用一個非引數化分類方法(參考 參考文件)來建立一個模型,預測新的戶外裝備影像的分類情況,訓練來自挑戰2的預處理過的128x128x3的裝備影像。所使用的演算法可以從scikit-learn庫中挑選現有的非引數化演算法來做分類。向教練展示所提供的測試資料集的精確度,並且精確度分數需要超過80%。
dir_data ="data/preprocess_images/" equipments = ['axes', 'boots', 'carabiners', 'crampons', 'gloves', 'hardshell_jackets', 'harnesses', 'helmets', 'insulated_jackets', 'pulleys', 'rope', 'tents'] train_data = [] y = [] import os from PIL import Image for equip_name in equipments: dir_equip = dir_data + equip_name for filename in os.listdir(dir_equip): if(filename.find('jpeg')!=-1): name = dir_equip + '/' + filename img = Image.open(name).convert('L') train_data.append(list(img.getdata())) y.append(equip_name)
from sklearn import svm from sklearn.cross_validation import train_test_split train_X,test_X, train_y, test_y = train_test_split(train_data, y, test_size = 0.3, random_state = 0) from sklearn import neighbors from sklearn.metrics import precision_recall_fscore_support as score from sklearn.metrics import precision_score,recall_score clf_knn = neighbors.KNeighborsClassifier(algorithm='kd_tree') clf_knn.fit(train_X, train_y) y_pred = clf_knn.predict(test_X)
print(__doc__) import itertools import numpy as np import matplotlib.pyplot as plt from sklearn import svm, datasets from sklearn.model_selection import train_test_split from sklearn.metrics import confusion_matrix def plot_confusion_matrix(cm, classes, normalize=False, title='Confusion matrix', cmap=plt.cm.Blues): """ This function prints and plots the confusion matrix. Normalization can be applied by setting `normalize=True`. """ if normalize: cm = cm.astype('float') / cm.sum(axis=1)[:, np.newaxis] print("Normalized confusion matrix") else: print('Confusion matrix, without normalization') print(cm) plt.imshow(cm, interpolation='nearest', cmap=cmap) plt.title(title) plt.colorbar() tick_marks = np.arange(len(classes)) plt.xticks(tick_marks, classes, rotation=45) plt.yticks(tick_marks, classes) fmt = '.2f' if normalize else 'd' thresh = cm.max() / 2. for i, j in itertools.product(range(cm.shape[0]), range(cm.shape[1])): plt.text(j, i, format(cm[i, j], fmt), horizontalalignment="center", color="white" if cm[i, j] > thresh else "black") plt.tight_layout() plt.ylabel('True label') plt.xlabel('Predicted label') # Compute confusion matrix # cnf_matrix = confusion_matrix(y_test, y_pred) np.set_printoptions(precision=2) confusion_mat = confusion_matrix(test_y, y_pred, labels = equipments) # Plot non-normalized confusion matrix plt.figure(figsize=(10,10)) plot_confusion_matrix(confusion_mat, classes=equipments, title='Confusion matrix, without normalization') # Plot normalized confusion matrix plt.figure(figsize=(10,10)) plot_confusion_matrix(confusion_mat, classes=equipments, normalize=True, title='Normalized confusion matrix') plt.show()
因為要求精確度>0.8,sklearn中的很多演算法應該都能滿足,我選擇了準確度比較高的KNN來建模,應該足夠用了
算一下presion和recall,輕鬆超越0.8
題目4
挑戰完成標準,使用深度學習模型,如CNN分析複雜資料
團隊將在Jupyter Notebook上為教練執行一個程式碼單元,展示模型的準確度為90%或更高
準確度如果要>0.9,sklearn中的機器學習演算法就很難達到了,關鍵時刻只能上CNN
import matplotlib.pyplot as plt from PIL import ImageColor from matplotlib.pyplot import imshow from PIL import Image from pylab import * dir_data ="data/preprocess_images/" equipments = ['axes', 'boots', 'carabiners', 'crampons', 'gloves', 'hardshell_jackets', 'harnesses', 'helmets', 'insulated_jackets', 'pulleys', 'rope', 'tents'] train_data = [] y = [] import os from PIL import Image i=0 for equip_name in equipments: dir_equip = dir_data + equip_name for filename in os.listdir(dir_equip): if(filename.find('jpeg')!=-1): name = dir_equip + '/' + filename img = Image.open(name).convert('L') train_data.append(array(img).tolist()) y.append(i) i += 1 train_data = np.asarray(train_data)
from sklearn import svm from sklearn.cross_validation import train_test_split import numpy as np import keras num_classes=12 img_rows=128 img_cols=128 train_X, test_X, train_y, test_y = train_test_split(train_data, y, test_size = 0.3, random_state = 0) train_X = train_X.reshape(train_X.shape[0], img_rows, img_cols, 1) test_X = test_X.reshape(test_X.shape[0], img_rows, img_cols, 1) train_X = train_X.astype('float32') test_X = test_X.astype('float32') train_X /= 255 test_X /= 255 print('x_train shape:', train_X.shape) print(train_X.shape[0], 'train samples') print(test_X.shape[0], 'test samples') # convert class vectors to binary class matrices train_y = keras.utils.to_categorical(train_y, num_classes) test_y = keras.utils.to_categorical(test_y, num_classes)
from keras.layers import Dense, Activation, Convolution2D, MaxPooling2D, Flatten from keras.models import Sequential from keras.layers import Convolution2D,MaxPooling2D, Conv2D import keras model = Sequential() model.add(Conv2D(32, kernel_size=(3, 3), activation='relu', input_shape=(128, 128, 1))) model.add(MaxPooling2D(pool_size=(2, 2))) model.add(Conv2D(64, (3, 3), activation='relu')) model.add(MaxPooling2D(pool_size=(2, 2))) # model.add(Dropout(0.25)) model.add(Flatten()) model.add(Dense(128, activation='relu')) # model.add(Dropout(0.5)) model.add(Dense(12, activation='softmax')) model.compile(loss=keras.losses.categorical_crossentropy, optimizer=keras.optimizers.Adadelta(), metrics=['accuracy']) model.fit(train_X, train_y, batch_size=128, epochs=50, verbose=1, validation_data=(test_X, test_y)) score = model.evaluate(test_X, test_y, verbose=0) print('Test loss:', score[0]) print('Test accuracy:', score[1])
CNN的混淆矩陣比KNN的好了不少
訓練了好多次,不斷調整各個卷積層和引數,終於達到了一個比較好的效果~~
題目6
使用深度學習框架,基於一個常用的模型,比如Faster R-CNN,訓練一個目標檢測的模型。這個模型需要能夠檢測並且使用方框框出圖片中出現的每一個頭盔。
這道題目首先要自己標註樣本,幾百張影像標註完累的半死。。。這裡我們使用VOTT來標註,它會自動生成一個樣本描述檔案,很方便。Faster R-CNN的程式我們參考了git上的一個紅細胞檢測的專案,https://github.com/THULiusj/CosmicadDetection-Keras-Tensorflow-FasterRCNN,程式碼非常多就不貼了
最後來一張效果圖
本文資料集和VOTT工具 連結:
https://pan.baidu.com/s/1FFw0PLJrrOhwR6J1HexPJA
提取碼 s242