適用於高性價比智慧型手機的實時影像處理算法

Abstract

在現有的高端智慧型手機中，因為圖形處理器強大的運算能力，多被用以实现影像增強算法。然而，高成本的圖形處理器同時造成高昂的手機價格，致使一些來自新興市場的使用者無法負擔。因此，在高性價比智慧型手機中，低成本的通用驅動芯片取代了圖形處理器用以實現影像增強算法。另外，常用於儲存影像訊息的高速幀緩衝器也因為其高昂的成本而被移除。考慮到通用驅動芯片的運算能力較低，本論文提出一種低運算複雜度且無需幀緩存器的影像增強算法，并將該算法應用於高性價比智慧手機以提高其市場競爭力。 本文中所提出的演算法包含四大主軸：白平衡校正、動態範圍增強、飽和度增強和美顏模式。白平衡校正可以校正影像色溫，且相比傳統方法，本文中提出的方法不會因影像存在大色塊而發生色偏。動態範圍增強算法可以使影像顯現出更多細節或增強其對比度。此外，飽和度增強不僅可以提供更鮮艷的自然風景影像，而且當影像中存在人物時，膚色區域亦可不受飽和度的提升的影響。最後，美顏模式可以為影像中的人物提供東方人更喜愛的白皙紅潤之肌膚。 考慮到通用型驅動芯片的運算能力有限，算法中的映射公式均使用查表法實現。因為未使用幀緩存器，影像訊息只能逐行分析。另外，在人因實驗中，受測者被要求對不同類型的影像進行主觀評分（5：很好 4：好 3：普通 2：差 1：極差）。調整后影像的分數相較於原影像提升了30%或者達到4分以上，證明了此算法可以提升影像品質。最後，對於1080×1920的影像進行模擬，此演算法的處理時間為20 ms，達到了實時處理的要求。

In existing high-end smartphones, the image-enhancement algorithm that makes displayed images clear, vivid, and natural is usually realized by a graphics processing unit (GPU) with strong computing power. However, the high cost of GPUs makes high-end smartphones unaffordable to consumers in emerging markets. Therefore, instead of an expensive GPU, a low-cost general driver integrated circuit (IC) is desired to implement the image-enhancement algorithm in low-cost smartphones. Moreover, the high-speed frame buffer that is commonly used to store images in the GPU-based algorithm is a high-cost component that is expected to be removed. Therefore, an image-enhancement algorithm with much lower computational complexity and no need for a frame buffer is proposed in this thesis to improve the competitiveness of low-cost smartphones. This algorithm must be able to be processed in real time so that it can be applied to videos. The proposed image-enhancement algorithm consists of four main parts: white balance correction, dynamic range enhancement, saturation enhancement, and make-up mode. The white balance correction method can adjust the image to make it appear as if it was taken under a canonical light source. Compared with the conventional method, in the white balance correction method, the correction performance is less affected by large-area objects with uniform color. Subsequently, a dynamic range enhancement method is implemented to reproduce as many image details as possible or to maximize the image contrast. The saturation-enhancement method can increase images’ vividness without affecting the skin color. Moreover, the make-up mode method can produce portraits with paler and rosier complexions. A subjective study was also conducted in which every image is assigned an opinion score (5: Excellent, 4: Good, 3: Fair, 2: Poor and 1: Bad). The mean opinion score (MOS) of the adjusted images is increased by at least 30%, while the original images have score lower than 3 or above 4, indicating that the image quality was significantly enhanced by the proposed algorithm. Given the limited computing power of a general driver IC, the mapping function in each method is implemented using look-up tables. To further reduce the cost, images are analyzed line by line so that there is no need to use a frame buffer to store the image signals. Finally, the processing time is verified to be only 20 ms for a full high-definition (FHD) image using a post-layout simulation, demonstrating that the proposed algorithm can be executed in real time.