相機模組影像品質評價

Abstract

在監視相機中，有一種相機規格名叫視覺解析度，普遍上利用電視線數量做為評鑑相機視覺解析度的好壞。傳統上評鑑相機電視線的方法分為兩個步驟，一是利用斜邊刀口來計算相機模組的空間頻率響應，二是採用固定的敏感度比率與空間頻率響應的交點即是相機模組的電視線數量。然而此方法計算得到的電視線數量與人眼所評鑑出來的電視線的數量有約一百條的誤差，事實上此誤差的原因來自於人眼的視覺系統特性會隨著空間頻率的不同而改變。 在本論文中，人眼的視覺特性可以由對比敏感度曲線來呈現，接著同樣是利用斜邊刀口來計算相機模組的空間頻率響應，而兩曲線的交點即是相機模組的電視線數量。在實驗中，採用八台720P的相機模組及七個不同的對焦狀態來做為日間模式的實驗樣本；採用四台720P相機模組及七個不同的對焦狀態作為夜間模式的實驗樣本；請十個檢測員來評鑑相機模組在不同對焦狀態下的電視線數量。用提議的方法在日間模式中，不同對焦狀態下所得到的電視線數量與其對應檢測員所評鑑出的電視線數量的平均絕對值誤差僅有25.6 (與720P規格相比，約占3.6%)，而用傳統方法所計算出來的平均絕對值誤差則為66.6 (約9.2%)。在夜間模式中，提議的方法所得到的平均絕對值誤差為26.5 (約3.7%)，而傳統方法所計算出來的平均絕對誤差是94.3 (約13%)。 相機模組經常使用銳利化後製來增加其視覺解析度。我們相信提議的方法在銳利化的相機模組中仍然適用，因此在實驗中日間模式及夜間模式皆採用四台720P相機模組及六個不同對焦狀態。在日模式中，提議的方法在不同對焦狀態下所得到的電視線數量與其對應檢測員所評鑑出的電視線數量其平均絕對值誤差為15.52 (約2.1%)；在夜間模式中，用提議的方法其平均絕對值誤差為22.7 (約3.2%)。

Visual resolution is an important specification of a surveillance camera and it is usually evaluated by television line (TVL). The conventional evaluation method adopts slanted-edge spatial frequency response and uses a fixed decision contrast-ratio to figure out the TVL. However, this method brings about a considerable error with respect to results of manual evaluation since the perceptibility of human vision system varies from spatial frequency. Therefore, in this paper, contrast sensitivity function characterizing human vision system and slanted-edge spatial frequency response are combined to solve TVL, and a systematic calculation method is proposed. By adopting eight 720P camera modules working in day mode and four 720P modules working in night mode, mean absolute error between TVL calculated by proposed method and that judged by ten inspectors is as low as 25.6 (3.6% of 720P) while the error of the convention method is 66.6 (9,2% of 720P) in day mode; The mean absolute error in night mode between proposed method and that judged by inspectors is 26.5 (3.7% of 720P) while the error of the conventional method is 94.3 (as 13.0% of 720P). Besides, image sharpening is widely used in surveillance cameras in order to upgrade visual resolution of image. From previous result of image non-sharpening, we believe that the proposed method which considers contrast sensitivity which varies from spatial frequency also reflects improvement after image sharpening. However, the conventional method using fixed decision of contrast ratio can’t accurately match human vision system. Hence, by adopting four 720P camera modules working in day mode and four 720P modules working in night mode. The evaluation of image sharpening with proposed method was done and mean absolute error is 15.52 (2.1% of 720P) in day mode and 22.7 (as 3.2% of 720P) in night mode.