基於學習之整合式超解析度影像

Abstract

高解析度影像技巧隨著科技進步而不斷地發展，所對應的硬體規格也隨之提升。使用者需要更精確的影像擷取感測器與更大的儲存記憶體來滿足需求，這代價是十分昂貴的。為了降低取得高解析度影像的成本，利用「超解析度重建技術」將低解析度影像轉換成高解析度影像，或是將較小的影像提升取樣至較大影像的技術不斷地被研發與改進。 在這本篇論文中，我們提出一個基於學習之整合式超解析度影像方法。此方法主要是利用資料庫中高解析度片塊與低解析度片塊相對應的關係所建立出來的模型，去預測估計出低解析度輸入影像中所缺少的細節部分。我們的系統分成兩大部分：訓練階段以及合成階段。在訓練階段，我們會建立一個資料庫；在合成階段，我們會先取得適合的資料並建立自我相似模型然後更新資料庫。接著根據影像片塊的性質去選擇相對應的超解析度演算法，再利用反投影技巧去滿足全域重建限制，最後強化超解析度影像的細節以得到高解析度影像。 相較於現有以學習為基礎的超解析度方法，我們的方法是非常有效率的，而且大大地提高了影像的品質，不僅具有銳利的邊緣以及豐富的細節。

Nowadays, the requirement for image resolution increases fiercely. However, the cost of high resolution images obtained from those modern devices is usually expensive, and it is not easy for people to afford. Therefore, the techniques called “super-resolution” enhancing the low resolution image to higher one are quite important. In recent decades, many researches were dedicated in this field and plenty of algorithms were proposed. In this thesis, we present an integrated learning-based super-resolution. Learning-based super-resolution techniques model the co-occurrence patterns between the high and low resolution patches of example images to estimate the missing details for low resolution input. Our system has two parts: training phase and synthesis phase. In the training phase, we construct a database. And in synthesis phase, we retrieve some suitable data and build multi-scale self-similarity model to update the database. We choose corresponding super-resolution algorithms based on different content, and we use back-projection to enforce global reconstruction constraint, and then enhance details of the super-resolved image. Comparing to existing learning-based approaches, our proposed method significantly improves image quality, and the produced super-resolution images have sharp edges and rich details; moreover, the algorithm is very efficient.