- 學位論文
以類神經網路為最佳化途徑來完成影像強化
The Optimization Approach of Image Enhancement by Neural Network
摘要
由於科技的進步，造就了數位影像商品的普及，例如：數位相機，照相手機，數位電視等。所以如何把所擷取到的影像來優化，漸漸受到人們的重視。雖然數位影像的感應器越做越精良，但是所收集到的影像資訊仍然與人眼所看到的會有些差距。所以必須靠著分析影像的組成，針對不同的本質，來做優化。影像的組成大致分成亮度與色度。擁有好的亮度分佈的影像，所包含的色度也就越充實，人眼看起來當然就感覺舒服； 若影像亮度分配不均，過亮或過暗，也就把色度的資訊隱藏起來，就好比人眼在大太陽光下或是在強烈的雪地反光下，看到的幾乎都是白色，或是在缺乏光源的情況下，所看到的幾乎都是灰色的資訊一樣，人眼看起來一定不舒服。如何找出每一張影像最佳的亮度分佈，即是本論文所討論的重點。我們所提出的方法是以類神經網路來做為最佳化的途徑，以梯度為概念，使的每一次的疊代所找到的解均朝著亮度最佳化的方向前進，若找尋方向可能有錯，也可以即時的修正，最後收斂在最佳解。我們也實驗了基因演算法與粒子群優演算法來作為找尋最佳解的途徑，因為這兩種演算法的核心其實大同小異，只是粒子群優演算法的過程比較簡潔。經過我們的實做，發現類神經網路所找出來的結果，比基因演算法及粒子群優演算法所找到的還要更好，並且疊代的次數也比較少，性能更加優越。
並列摘要
As the growth of technology, the product about digital image becomes more and more popular. People gradually pay much attention about the enhancement of image captured by the digital equipment. Though the sensor are well-designed, the captured image is still different from the sight of people in some way, and the image with bad luminance distribution will result in the disappearance of the chrominance information. In order to evaluate the image quality, we quantize it by a function. And then, the problem is how to optimize the system. If the analytical form of the system exists, the extreme value can be easily derived from the partial derivative of the function. If not, we have to take advantage of some optimization algorithm to find the extreme value. Traditional random-based optimization methods such as genetic algorithm and particle swarm optimization do not find the best score efficiently. We utilize the neural network to model the nonlinear relationship between the input and output and gradient information which will adaptively guide the system to the state of optimization in less iterations than GA and PSO. The performance of the proposed work is better than the two algorithms mentioned before.