受限於光學鏡頭的場景深度限制,物體只有在景深限制內被拍攝時會清晰的成像,因此,現有常見的數位相機不容易拍攝出一張所有物體都清晰的影像,而多重對焦影像融合技術可以解決這個問題,將多張具有相同場景、不同深度限制的影像融合成出一張所有物體相對清晰的影像。在本論文中,提出一個新的影像融合架構,利用特徵擷取以及使用 SVM 作分類的多重對焦影像融合方法。首先,擷取出影像的四個特徵,並將影像區分成三個區域,分別是對焦、離焦、以及不確定區域,將三個區域的特徵分別帶入 SVM 模組,利用前兩區域的特徵來對不確定區域作分類,再來依分類的結果產生初始的權重,最後以 sum-modified-Laplacian (SML) 修正初始權重,再進行影像融合取得合成影像。根據本研究的實驗結果顯示,本論文所提出的研究方法優於比較的四種現有方法。
In terms of the finite depth-of-field of optical lenses, it is difficult to capture an image with all objects clear by common cameras. Only objects within the depth-of-field are captured in focus and sharp while others are defocus and blurred. Multi-focus image fusion technique is aimed to integrate multiple images with difference focused depth of the same scene to generate an image with every object are focused. In this study, a multi-focus image fusion based on feature extraction and SVM classification is proposed. Within the proposed approach, four various features, namely, intensity, gradient, partition, and saliency, are extracted. Then, each source image is segmented into three categories: focus, defocus, and undefined regions. Pixels belong to focus and defocus regions are regard as the input of SVM training. After performing SVM training and classification, initial weighting maps are estimated and a focus measure, sum-modified-Laplacian (SML), is applied to refined the initial weighting maps and establish the final weighting maps. Base on the experimental results obtained in this study, the performance of the proposed approach is better than those of four comparison approaches.