| 摘要 |
依據文獻指出 3D
立體影像之交錯殘影(Crosstalk)是影響影像品質和視覺疲勞最大的主要原因，常造成觀看者心理與生理的負擔。目前 3D
立體光學設計乃憑藉著設計者過去先前累積的經驗或由光學軟體模擬所得到的結果，最後選擇了較保守的設計參數，這樣不僅不容易掌握設計參數品質，無形中也花費了許多時間與成本。因此，在要求設計、製造效率與加工品質前提下，若能事先掌握光學模擬設計的預測模型，以及解決設計參數的最佳化問題，確有其必要性。本研究為建立裸眼
3D 立體影像之 Crosstalk 品質條件下，光學模擬設計參數最佳化模組。其中應用類神經網路(Artificial
Neural Network ; ANN)之倒傳遞網路(Back-Propagation Neural Network ;
BPNN)，建構裸眼 3D 立體影像之
Crosstalk品質預測模型，以模擬的設計數據作為輸入參數，以模擬後落於兩眼的輝度值作為輸出目標值，並利用模擬實驗的量測結果來進行網路訓練與測試後，預測模型的推論值與目標值相比較，誤差值小於
1%，模型驗證誤差也小於 1%。將建立類神經預測模型之預測值做為田口方法的觀測值，利用田口方法之望小特性 SN比求出最佳因子組合為
A3B1C1D1，確認實驗數值為 48.024 落在信賴區間[32.926,66.402]之內，表示確認實驗落於
95%信賴區間內，故本研究之類神經預測模型是可信賴的。利用最佳因子組合 A3B1C1D1 為設計依據，以微柱狀透鏡(Lenticular
Lens)為例，驗證結合奈米壓印技術與 LIGA LIKE 製程，微柱狀透鏡(Lenticular Lens)曲率半徑為 228
μm、高度 18μm，使用共焦 3D 雷射顯微鏡量測金屬模仁、鎳(Ni)模仁、PDMS
母模仁與製作薄膜成品曲率半徑與間距，量測結果發現製作完成微柱狀透鏡之幾何形狀與設計值非常接近，幾何尺寸誤差都在
5%以下，表面粗糙度達到光學鏡面等級(Ra 為均為 10 nm 以下)，複製率達到
95%，結果驗證此製程可以製作出單一曲率、無間隙、外形長扁形且高精度之微透鏡陣列。可見，設計模擬參數最佳化預測模組能有效且快速求得最佳設計參數組合，並可降低製程II加工後交錯殘影(Crosstalk)品質的要求並獲得最大效益。
關鍵字： 類神經網路倒傳遞類神經網路，交錯殘影，奈米壓印，LIGA-Like製程
Crosstalk is a major factor
affecting the quality of 3D stereoscopic images. High levels of
crosstalk may cause visual fatigue and mental and physical
discomforts. Current 3D stereoscopic optics is created based on
past designs and simulation results of optical software. Using
conservative design parameters not only makes the design quality
difficult to control but also wastes time and costs. Creating a
predictive optical simulation model is necessary to solve the
optimization problem in parameter designs, thereby ensuring
image quality and manufacturing efficiency. This research aims
to build an optimization module for optics design parameters to
improve the quality of naked-eye 3D stereoscopic images. The
proposed module applies back-propagation of artificial neural
network to construct a model for crosstalk prediction in
naked-eye 3D stereoscopic images. The performance of the model
was evaluated by using simulated design parameters as input
parameters and binocular luminance as target output value.
Simulation results were then utilized for network training and
testing. The difference of the deduction value of the prediction
model and the target value is less than 1%, and the model
validation is also less than
1%. Take the predictive value of the artificial neural
prediction model as the observed value of the Taguchi method; we
can determine the combination of the optimal factors, which is
A3B1C1D1, by using the signal-to-noise ratio from the
smaller-the-better response of the Taguchi method. The
confirmation experimental value is 48.024, which falls within
the confidence interval[32.926, 66.402], thereby denoting that
the experiment lies inside the 95% confidence interval.
Therefore, the artificial neural prediction model is reliable.
This study utilized the optimal factor combination A3B1C1D1 as
the design basis IV and the microlenticular lens as an example
to verify the combination of nanoimprint technology with LIGA-like
process. The curvature radius and height of the microlenticular
lens are 228 and 18 μm, respectively. Confocal 3D laser
microscopy was used to measure the curvature radius and spacing
of thin films made with metal, Ni, and polydimethylsiloxane
molds. Results show that the geometric shape of the produced
microlenticular lens is close to the design values, in which the
geometric error is less than 5%. The surface roughness reaches
the optical specular level (Ra of less than 10 nm), and the
replication rate reaches 95%. Therefore, this process can
produce single curvature, gap-free, long, flat, and
high-precision microlens arrays. The optimization module can
lower the post-processing crosstalk quality requirement
effectively. The optimization module for optics design
parameters can effectively and quickly obtain the best design
parameter combination and reduce the quality requirement of
interlaced blur image after processing.
Key words: Artificial Neural
Network、Back-PropagationNetwork、Crosstalk、Nano-imprint、LIGA-Like
|
