이미지에서 특징점을 추출한다.
추출한 특징점의 설명자를 클러스터링 한다.
이미지의 특징점들의 클러스터를 하나의 단어라고 생각하고 count vector를 생성한다.
생성한 count vector를 이용하여, 분류 모델을 생성한다.
사용할 데이터 셋 - 손글씨 데이터
1 2 3 4
import cv2from keras.datasets import fashion_mnist(x_train, y_train), (x_test, y_test) = fashion_mnist.load_data()
SIFT 특징점 추줄
1 2 3 4 5 6 7
sift = cv2.xfeatures2d.SIFT_create(100 ) keypoints = [] descriptors = [] for img in x_test: kps, des = sift.detectAndCompute(img, None ) keypoints.append(kps) descriptors.append(des)
1 2 3 4 5
df = pd.DataFrame() df["kps" ] = keypoints df["des" ] = descriptors df["label" ] = y_test df.tail()
kps
des
label
9995
[<KeyPoint 0x7fdfce4031e0>, <KeyPoint 0x7fdfce...
[[0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 43.0...
9
9996
[<KeyPoint 0x7fdfce4032d0>, <KeyPoint 0x7fdfce...
[[2.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 16.0...
1
9997
[<KeyPoint 0x7fdfce403390>, <KeyPoint 0x7fdfce...
[[19.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 6.0, 92....
8
9998
[<KeyPoint 0x7fdfce403450>, <KeyPoint 0x7fdfce...
[[32.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 2.0, 191...
1
9999
[<KeyPoint 0x7fdfce403510>, <KeyPoint 0x7fdfce...
[[1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 2.0,...
5
이미지에서 찾은 특징점과 그 설명자(descriptor)
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35
plt.figure(figsize=(15 , 9 )) plt.subplot(1 , 2 , 1 ) plt.imshow(x_train[3 ], cmap='gray' ) pts_x = [kp.pt[0 ] for kp in df.kps[3 ]] pts_y = [kp.pt[1 ] for kp in df.kps[3 ]] plt.scatter(pts_x[:4 ], pts_y[:4 ], s=200 , c='r' ) plt.axis('off' ) des = df.des[418 ] for i in range(4 ): plt.subplot(4 , 2 , (i+1 )*2 ) plt.bar(np.arange(len(des[i])), des[i]) plt.xticks(range(4 , len(des[i]), 8 )) plt.title("숫자 4의 {}번째특징점의 설명자" .format(i)) plt.tight_layout() plt.show() plt.figure(figsize=(15 , 9 )) plt.subplot(1 , 2 , 1 ) plt.imshow(x_train[516 ], cmap='gray' ) pts_x = [kp.pt[0 ] for kp in df.kps[10 ]] pts_y = [kp.pt[1 ] for kp in df.kps[10 ]] plt.scatter(pts_x[:4 ], pts_y[:4 ], s=200 , c='r' ) plt.axis('off' ) des = df.des[101 ] for i in range(4 ): plt.subplot(4 , 2 , (i+1 )*2 ) plt.bar(np.arange(len(des[i])), des[i]) plt.xticks(range(0 , len(des[i]), 32 )) plt.title("숫자 3의 {}번째특징점의 설명자" .format(i)) plt.tight_layout() plt.show()
1 2 3 4 5 6 7 8
df = df.dropna() df = df.reset_index(drop=True ) X = [] for des in df.des: for d in des: X.append(d) X = np.asarray(X)
1 2 3 4
from sklearn.cluster import KMeansmodel_clu = KMeans(n_clusters=8 , init="k-means++" ).fit(X) clu = model_clu.predict(X)
클러스터링 결과의 시각화
1 2 3 4 5 6
from sklearn.manifold import TSNEX_embedded = TSNE().fit_transform(X[:1000 ]) plt.scatter(X_embedded[:, 0 ], X_embedded[:, 1 ], c=clu[:1000 ], cmap="tab20" ) plt.title("클러스터링 결과 시각화" ) plt.show()
설명자의 클러스터로 Count Vector 생성
1 2 3
def (des) : clu_nums = model_clu.predict(des.reshape(-1 , 128 )) return np.bincount(clu_nums, minlength=20 )
이미지에서 추출한 특징점들의 설명자를 클러스터 카운트 벡터로 변환
1 2 3 4
X = [] for des in df.des: X.append(vbow_vectorizer(des)) y = df.label
[array([0, 0, 1, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]),
array([0, 2, 0, 1, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]),
array([0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]),
array([0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]),
array([0, 1, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0])]
분류 모델 생성
1 2 3
from sklearn.naive_bayes import MultinomialNBmodel_cls = MultinomialNB().fit(np.asarray(X), y)
분류기 성능
1 2 3 4 5
from sklearn.metrics import classification_reporty_true = y y_pred = model_cls.predict(X) print(classification_report(y_true, y_pred))
precision recall f1-score support
0 0.26 0.44 0.33 944
1 0.32 0.41 0.36 628
2 0.22 0.04 0.07 954
3 0.22 0.27 0.24 923
4 0.21 0.13 0.16 973
5 0.39 0.56 0.46 966
6 0.19 0.09 0.12 932
7 0.32 0.38 0.35 965
8 0.32 0.25 0.28 958
9 0.46 0.59 0.52 987
avg / total 0.29 0.31 0.29 9230
1 2 3 4
def test (img) : kps, des = sift.detectAndCompute(img, None ) vbow = vbow_vectorizer(des) return model_cls.predict(vbow.reshape(-1 ,20 ))
1 2 3 4
print(test(x_train[11 ])) plt.imshow(x_train[11 ], cmap="gray" ) plt.axis('off' ) plt.show()
[5]
1 2 3 4 5 6 7 8
from sklearn.metrics.pairwise import cosine_similaritydef check_similarity (img1, img2) : kps1, des1 = sift.detectAndCompute(img1, None ) kps2, des2 = sift.detectAndCompute(img2, None ) vbow1 = vbow_vectorizer(des1) vbow2 = vbow_vectorizer(des2) return cosine_similarity(vbow1.reshape(-1 , 20 ), vbow2.reshape(-1 , 20 ))
1 2 3 4 5 6 7 8 9 10
plt.figure() print("두 이미지 간의 유사도 = {}" .format(check_similarity(x_train[11 ], x_train[1 ]))) plt.subplot(1 ,2 ,1 ) plt.imshow(x_train[11 ], cmap="gray" ) plt.axis('off' ) plt.subplot(1 ,2 ,2 ) plt.imshow(x_train[1 ], cmap="gray" ) plt.axis('off' ) plt.tight_layout() plt.show()
두 이미지 간의 유사도 = [[0.65659915]]
