| 摘要: | 本研究以實驗室現有的14頻道燈箱為基礎，設計一款可以藉由輸入光參數後，快速得到符合輸入光參數光譜的光譜生成器。利用基因演算法來生成不同照度的數據集，使用數據集來訓練神經網路，透過不同超參數組合來評估表現良好的模型，並將該模型設計成光譜生成器。本研究的光譜生成器分成無拓展輸入與拓展輸入兩種，無拓展輸入情況下，光譜生成時間為0.01至0.02秒，而拓展輸入時，光譜生成時間為0.19秒，雖然時間比無拓展輸入久但可以預測較大的範圍。
透過基因演算法生成相關色溫(Correlated Color Temperature, CCT)為3000 K、4000 K、5000 K與6500 K，照度(Illuminance, Ev)為500 lux、600 lux、700 lux、750 lux、800 lux、900 lux以及1000 lux的數據集，其中會包含14頻道權重、CCT、色偏差值(Delta u-v, Duv)、顏色保真度指數(Color Fidelity Index, Rf)以及晝夜節律刺激值(Circadian Stimulus, CS)等光參數。
在訓練神經網路的部分，會先將各照度的數據集拆分成訓練集、驗證集以及測試集，再依照實驗規劃將不同照度的訓練、驗證與測試集組合成新的訓練、驗證與測試集，並對訓練集與驗證集進行二維過採樣，使用訓練集與驗證集進行訓練。為了使神經網路的訓練時間以及訓練誤差達到最小，本研究比較了不同照度的數據集組合以及不同的超參數組合，來使神經網路可以用最少的數據以及最短的訓練時間來達到最好的效果。
研究結果顯示，當改變Ev時，CS值也會跟著變大，但Ev與CS並不為線性關係，Ev越大時，CS值的最大值與最小值的範圍將會變小。CS值也會受到藍色-黃色機制(b-y)的影響，使4000 K的CS值低於其他三種色溫。訓練神經網路時，適當的正規化值(L2 regularization)可以有效的降低神經網路的預測誤差。最終光譜生成器將可預測不同照度下不同照明參數的光譜，只要是位於數據集內的光參數，光譜皆可以順利被預測生成。
;In this study, we design a spectrum generator based on the 14-channel illuminator that can quickly produce spectra matching the input lighting parameters. A genetic algorithm is used to generate datasets of different illuminances, which are then used to train neural networks. Various hyperparameter combinations are evaluated to develop a well-performing model, which is then used to create the spectrum generator.
The spectrum generator in this study is divided into two types: one with non-extended input and one with extended input. For non-extended input, the spectrum generation time is 0.01 to 0.02 seconds, while for extended input, it is around 0.19 seconds. Although the extended input takes a longer time, it allows for the prediction of a broader range.
Using a genetic algorithm, datasets were generated with Correlated Color Temperatures (CCT) of 3000 K, 4000 K, 5000 K, and 6500 K, and Illuminance (Ev) of 500 lux, 600 lux, 700 lux, 750 lux, 800 lux, 900 lux, and 1000 lux. These datasets include lighting parameters such as 14-channel weights, CCT, Delta u-v (Duv), Color Fidelity Index (Rf), and Circadian Stimulus (CS).
For the neural network training, the dataset for each illuminance is first split into training, validation, and test sets. According to the experimental plan, these sets are then combined into new training, validation, and test sets with different illuminances. The training and validation sets undergo two-dimensional oversampling and are used for training. To minimize the training time and error of the neural network, this study compares different combinations of illuminance datasets and various hyperparameter configurations. The goal is to achieve the best performance with the least amount of data and the shortest training time.
The research results indicate that as the Ev increases, the CS value also increases. However, the relationship between Ev and CS is not linear. As Ev becomes larger, the range between the maximum and minimum CS values decreases. The CS value is also influenced by the blue versus yellow color mechanism (b-y), resulting in lower CS values for 4000 K compared to the other three CCTs. When training the neural network, an appropriate L2 regularization value effectively reduces the prediction error. Ultimately, the spectrum generator can predict the spectrum of different lighting parameters under various illuminances, and as long as the lighting parameters are within the dataset, the spectra can be successfully generated. |
