使用類神經網路偵測影像色偏及模糊系統修正影像色偏之研究

Abstract

本論文提出的系統主要分為兩個部份：系統的第一個部份為使用類神經網路判斷影像中的色彩為無色偏、真實色偏(real cast)、或是含有本質色偏(intrinsic cast)。我們分別從350張訓練影像中，逐張取得13個具有代表性的統計參數，再將這些參數做為訓練倒傳遞神經網路(back-propagation neural network)的輸入特徵向量，並提供一組目標向量並使用監控式的學習法則來訓練類神經網路。系統的第二部份為使用模糊系統建構一個白平衡演算法(white balance algorithm)以修正色偏。我們先將影像分成若干個方形區域，分別對每個方形區域求得模糊系統的輸出權重，再配合亮度影響權重，最後可得到整張影像的色彩修正權重。 色偏偵測最大的困難是在於如何能夠準確的分辨出影像中的本質色偏，目前使用的色偏偵測法為臨界值偵測法[12]及長條圖偵測法[13]，而本論文使用類神經網路是為了提高本質色偏的偵測正確率，偵測結果將與兩種方法作比較。模糊系統白平衡演算法是為了改善灰界理論法(gray world assumption)[1]及色彩標準差調整灰界理論法(standard deviation weighted gray world assumption)[21]的缺點而提出的新方法，將利用含有色偏的馬克貝斯色票圖(Macbeth color chart)與灰界理論法、最大參考白演算法(max white method)[2]及色彩標準差調整灰界理論法的輸出結果做比較。另外，並挑選具代表性的自然影像對模糊系統白平衡演算法做測試。經由實驗結果可看出，類神經網路偵測影像色偏法對偵測本質色偏的正確率高於長條圖偵測法。模糊系統白平衡演算法對色彩數量較少或是色彩標準差較小的影像可以得到不錯的色彩修正效果。

There are two main parts in this thesis. The first part of the system is a color cast detector using the neural network. In this stage, the test images can be classified as having no cast, real cast, or intrinsic cast (image presenting a cast due to a predominant color that must be preserved). We have a data set of 700 images downloaded from internet, or acquired using various digital cameras. Choose 350 images from the data set as the training images for the neural network, and the rest 350 images for testing. From each training image, we can acquire 13 statistical parameters, and let them as the input vectors of the neural network. We also provide the neural network with the corresponding target vectors, and the supervised training method is used to train the neural network. The second stage is the white balance algorithm. If the real cast is found by color cast detector, the white balance algorithm should be applied on the test image. The test image is divided into n blocks. For each block, the output weighting can be obtained by a fuzzy system and the luminance weighted value is also calculated. Finally, we can obtain the new amplifier gains of the R, G and B channel to remove the color cast. The difficulty of the cast detector is how to distinguish the intrinsic cast from the test images. Existing methods of cast detection are threshold method [12] and histogram method [13], the performance of neural network cast detector will compared with these two methods. The proposed white balance algorithm will compared with the gray world assumption [1], max white method [2], and standard deviation weighted gray world assumption [21] by experimenting on the Macbeth color charts with color cast. Some representative images which contents are nature scenes will be tested by the proposed method. The experimental results show that the performance of detecting the intrinsic cast inside the test images using neural network cast detector is better than the histogram method. For the images which have less number of colors or small standard deviation of colors, the proposed white balance algorithm can improve the quality of color correction.