白平衡：即白色的平衡，最早用於攝像領域技術，可以用來解決色彩還原和調處理的一系列問題。
網上參考別人python版白平衡的運算，索性自己優化了下代碼。
用numpy矩陣運算取代原有的通道運算，提高運行速度。
tip：用python圖像處理盡可能用numpy運算（分通道寫法會使處理速度變慢，特別是較大圖片），而在C++中的圖像處理需要分通道處理。
下面是優化過整理過的代碼：
def WhiteBlance(img, mode=1): """白平衡處理（默認為1均值、2完美反射、3灰度世界、4基於圖像分析的偏色檢測及顏色校正、5動態閾值）""" # 讀取圖像 b, g, r = cv2.split(img) # 均值變為三通道 h, w, c = img.shape if mode == 2: # 完美反射白平衡 ---- 依賴ratio值選取而且對亮度最大區域不是白色的圖像效果不佳。 output_img = img.copy() sum_ = np.double() + b + g + r hists, bins = np.histogram(sum_.flatten(), 766, [0, 766]) Y = 765 num, key = 0, 0 ratio = 0.01 while Y >= 0: num += hists[Y] if num > h * w * ratio / 100: key = Y break Y = Y - 1 sumkey = np.where(sum_ >= key) sum_b, sum_g, sum_r = np.sum(b[sumkey]), np.sum(g[sumkey]), np.sum(r[sumkey]) times = len(sumkey[0]) avg_b, avg_g, avg_r = sum_b / times, sum_g / times, sum_r / times maxvalue = float(np.max(output_img)) output_img[:, :, 0] = output_img[:, :, 0] * maxvalue / int(avg_b) output_img[:, :, 1] = output_img[:, :, 1] * maxvalue / int(avg_g) output_img[:, :, 2] = output_img[:, :, 2] * maxvalue / int(avg_r) elif mode == 3: # 灰度世界假設 b_avg, g_avg, r_avg = cv2.mean(b)[0], cv2.mean(g)[0], cv2.mean(r)[0] # 需要調整的RGB分量的增益 k = (b_avg + g_avg + r_avg) / 3 kb, kg, kr = k / b_avg, k / g_avg, k / r_avg ba, ga, ra = b * kb, g * kg, r * kr output_img = cv2.merge([ba, ga, ra]) elif mode == 4: # 基於圖像分析的偏色檢測及顏色校正 I_b_2, I_r_2 = np.double(b) ** 2, np.double(r) ** 2 sum_I_b_2, sum_I_r_2 = np.sum(I_b_2), np.sum(I_r_2) sum_I_b, sum_I_g, sum_I_r = np.sum(b), np.sum(g), np.sum(r) max_I_b, max_I_g, max_I_r = np.max(b), np.max(g), np.max(r) max_I_b_2, max_I_r_2 = np.max(I_b_2), np.max(I_r_2) [u_b, v_b] = np.matmul(np.linalg.inv([[sum_I_b_2, sum_I_b], [max_I_b_2, max_I_b]]), [sum_I_g, max_I_g]) [u_r, v_r] = np.matmul(np.linalg.inv([[sum_I_r_2, sum_I_r], [max_I_r_2, max_I_r]]), [sum_I_g, max_I_g]) b0 = np.uint8(u_b * (np.double(b) ** 2) + v_b * b) r0 = np.uint8(u_r * (np.double(r) ** 2) + v_r * r) output_img = cv2.merge([b0, g, r0]) elif mode == 5: # 動態閾值算法 ---- 白點檢測和白點調整 # 只是白點檢測不是與完美反射算法相同的認為最亮的點為白點，而是通過另外的規則確定 def con_num(x): if x > 0: return 1 if x < 0: return -1 if x == 0: return 0 yuv_img = cv2.cvtColor(img, cv2.COLOR_BGR2YCrCb) # YUV空間 (y, u, v) = cv2.split(yuv_img) max_y = np.max(y.flatten()) sum_u, sum_v = np.sum(u), np.sum(v) avl_u, avl_v = sum_u / (h * w), sum_v / (h * w) du, dv = np.sum(np.abs(u - avl_u)), np.sum(np.abs(v - avl_v)) avl_du, avl_dv = du / (h * w), dv / (h * w) radio = 0.5 # 如果該值過大過小，色溫向兩極端發展 valuekey = np.where((np.abs(u - (avl_u + avl_du * con_num(avl_u))) < radio * avl_du) | (np.abs(v - (avl_v + avl_dv * con_num(avl_v))) < radio * avl_dv)) num_y, yhistogram = np.zeros((h, w)), np.zeros(256) num_y[valuekey] = np.uint8(y[valuekey]) yhistogram = np.bincount(np.uint8(num_y[valuekey].flatten()), minlength=256) ysum = len(valuekey[0]) Y = 255 num, key = 0, 0 while Y >= 0: num += yhistogram[Y] if num > 0.1 * ysum: # 取前10%的亮點為計算值，如果該值過大易過曝光，該值過小調整幅度小 key = Y break Y = Y - 1 sumkey = np.where(num_y > key) sum_b, sum_g, sum_r = np.sum(b[sumkey]), np.sum(g[sumkey]), np.sum(r[sumkey]) num_rgb = len(sumkey[0]) b0 = np.double(b) * int(max_y) / (sum_b / num_rgb) g0 = np.double(g) * int(max_y) / (sum_g / num_rgb) r0 = np.double(r) * int(max_y) / (sum_r / num_rgb) output_img = cv2.merge([b0, g0, r0]) else: # 默認均值 ---- 簡單的求均值白平衡法 b_avg, g_avg, r_avg = cv2.mean(b)[0], cv2.mean(g)[0], cv2.mean(r)[0] # 求各個通道所占增益 k = (b_avg + g_avg + r_avg) / 3 kb, kg, kr = k / b_avg, k / g_avg, k / r_avg b = cv2.addWeighted(src1=b, alpha=kb, src2=0, beta=0, gamma=0) g = cv2.addWeighted(src1=g, alpha=kg, src2=0, beta=0, gamma=0) r = cv2.addWeighted(src1=r, alpha=kr, src2=0, beta=0, gamma=0) output_img = cv2.merge([b, g, r]) output_img = np.uint8(np.clip(output_img, 0, 255)) return output_img
以下是白平衡效果圖
(原圖) (mode=1 均值) (mode=2 完美反射)
(mode=3 灰度世界) (mode=4 顏色校正) (mode=5 動態閾值)
處理的時間情況
1.均值：20ms
2.完美反射：20ms
3.灰度世界：20ms
4.基於圖像分析的偏色檢測及顏色校正：20ms
5.動態閾值算法:30ms
目前只針對網絡上的算法進行運算優化，待進行圖像分析及算法研究。