康除模糊：車牌影像除模糊

Abstract

在科技執法的領域中，監控攝像機捕捉到的模糊車牌圖像是一個常見且棘手的問題，這往往會阻礙交通違規行為的有效識別和及時處罰。由於車輛在行駛過程中速度快、環境光線變化大、攝像機角度和抖動等各種因素的影響，捕捉到的車牌圖像經常會變得模糊不清，從而使得傳統的圖像識別技術難以準確辨識車牌信息。因此，針對如何提升這些模糊車牌的可讀性，提供一個高效的去模糊解決方案成為了當前科技執法中一個亟需解決的問題。 我們針對目前的NAFNet (Nonlinear Activation Free Network)模型進行了架構上的改進，使其模型大小降低30%。這些改進不僅減少了模型的參數量，降低了硬體資源的需求，還顯著提高了模型的執行速度，從而更適合應用於實時圖像恢復的場景中。 在模型訓練過程中，我們採用了領域適應技術，旨在提升模型處理真實世界模糊影像的能力。這些技術幫助模型縮小了合成模糊數據與真實世界模糊影像之間的領域差距。實驗結果表明，與傳統的去模糊方法相比，我們的改進方法在多樣化的測試數據集上展現了更優異的復原效果。 總結來說，我們的研究為提升車牌圖像的可讀性提供了一個高效且實用的解決方案，通過改進模型架構和擴充數據集，成功應對了現實世界中多樣且不可預測的模糊挑戰。這不僅在技術上具有重要的創新意義，還為實際的交通科技執法應用提供了確實可行的技術支持。

In the realm of technological law enforcement, the challenge of blurred vehicle license plate images captured by surveillance cameras often impedes the efficient identification and penalization of traffic violations. This paper addresses the critical need for an effective deblurring solution that enhances the readability of license plates under such circumstances. The current State-Of-The-Art (SOTA) methods in image deblurring demonstrate limitations in handling the diverse and unpredictable conditions encountered in real-world scenarios. Our study modified the NAFNet model architecture, making it more compact and improving its execution speed, which is particularly beneficial for real-time image restoration. We have also incorporated domain adaptation techniques to enhance the model's ability to handle real-world blur scenarios. These techniques allowed the model to bridge the domain gap between synthetic blur data and real-world blur images. Experimental results demonstrate that, compared to traditional deblurring methods, our improved approach exhibits superior restoration performance across a diverse set of test datasets.