用例子講解圖像識別過程

圖像識別概念

圖像識別是識別圖像中物體的類別(屬於哪一個類)的任務。圖像識別通常被稱為Image Classification、Categorization、Clustering。
卷積神經網絡（CNN）出現之前，一般用HOG、SIFT、SURF等方法先從圖像中提取特征，然后通過特征確定物體的類別。

利用圖像直方圖實現簡單的圖像識別任務

算法流程：

1. 將訓練集中的圖像進行減色處理（圖像色彩量化：圖像色彩量化詳解）。RGB每個分量都只取4個值。
2. 創建訓練集減色圖像的直方圖。RGB圖像的直方圖中，B=[1,4]，G=[5,8]、R=[9,12]，此時bin=12，但是我還需要保留每張訓練圖所屬的類別，所以，bin=13。數據這樣存儲：database = np.zeros( (訓練數據數,13),dtype=np.int )。所有訓練數據的柱狀圖如下：
   
   database具有如下的形狀和內容(每一行最后一列是圖像所屬的類別)：
   
3. 將測試集圖像進行色彩量化，計算測試集圖像的直方圖與訓練集中每個直方圖的差，將差稱作特征向量。
4. 直方圖差異總和最小的訓練集中圖像的類別就是我們預測的待測圖像的類別。換句話說，待測圖像的類別與近色圖像一致。

實驗代碼：

import cv2
import numpy as np
import matplotlib.pyplot as plt
from glob import glob

# 色彩量化
def dic_color(img):
    img //= 63
    img = img * 64 + 32
    return img

# 創建 Database(db)
def get_DB():
    # get training image path
    train = glob("../dataset/train/*")
    train.sort()

    # prepare database
    db = np.zeros((len(train), 13), dtype=np.int32)

    # prepare path database
    pdb = []

    # each image
    for i, path in enumerate(train):
        # read image
        img = dic_color(cv2.imread(path))

        #get histogram
        for j in range(4):
            db[i, j] = len(np.where(img[..., 0] == (64 * j + 32))[0])
            db[i, j+4] = len(np.where(img[..., 1] == (64 * j + 32))[0])
            db[i, j+8] = len(np.where(img[..., 2] == (64 * j + 32))[0])

        # get class
        if 'akahara' in path:
            cls = 0
        elif 'madara' in path:
            cls = 1

        # store class label
        db[i, -1] = cls

        # store image path
        pdb.append(path)

    return db, pdb

# 判斷測試集中圖像類別
def test_DB(db, pdb):
    # get test image path
    test = glob("../dataset/test/*")
    test.sort()

    accurate_N = 0.

    # each image
    for path in test:
        # read image
        img = dic_color(cv2.imread(path))

        # get histogram
        hist = np.zeros(12, dtype=np.int32)
        for j in range(4):
            hist[j] = len(np.where(img[..., 0] == (64 * j + 32))[0])
            hist[j+4] = len(np.where(img[..., 1] == (64 * j + 32))[0])
            hist[j+8] = len(np.where(img[..., 2] == (64 * j + 32))[0])

        # get histogram difference
        difs = np.abs(db[:, :12] - hist)
        # axis=1表示以行為單位，求每一行的和
        difs = np.sum(difs, axis=1)

        # get argmin of difference
        pred_i = np.argmin(difs)

        # get prediction label
        pred = db[pred_i, -1]

        if pred == 0:
            pred_label = "akahara"
        elif pred == 1:
            pred_label = "madara"

        gt = "akahara" if "akahara" in path else "madara"

        if gt == pred_label:
            accurate_N += 1
        
        print(path, "is similar >>", pdb[pred_i], " Pred >>", pred_label)

    accuracy = accurate_N / len(test)
    print("Accuracy >>", accuracy, "({}/{})".format(int(accurate_N), len(test)))


if __name__ == '__main__':
	db, pdb = get_DB()
	test_DB(db, pdb)

實驗輸出（包含識別出的圖像的類別和識別准確率）：