應用於偵測視訊模糊物件的強健性類神經網路

Abstract

自21世紀電腦運算速度呈爆發性成長後，深度學習方法開始在許多領域進行推行與應用，其中針對物件偵測所設計的類神經網路以YOLO系列為大宗。然而在真實場景應用中，常需要面對影像因為錄製者本身的晃動、鏡頭變焦、物體移動，甚至是場景內的霧氣所帶來影像模糊的問題。本研究使用YOLOX作為基礎模型，改善了傳統YOLO模型應用於單幀模糊影像的表現，並與現有的多幀偵測方法YOLOV進行結合，實現在單幀與多幀情境下，均能進行穩定預測的強健性類神經網路。本研究改進了現有靜態物件偵測模型的前處理方法，額外加入了全局灰階高斯模糊影像進行訓練，並優化損失函數以契合模糊影像的預測需求，實現在性能改善的同時，又不需額外時間來進行預測的新模糊影像偵測模型，並兼具新網路模型應用於各情境及新模型的泛用性。

Since the explosive growth in computing speed in the 21st century, many applications of deep learning have been implemented across various fields. Currently, the neural networks designed for object detection primarily consist of the YOLO series. However, in real-world applications, images often face challenges such as motion blur from the recorder''s movement, camera zoom, object movement, or even image blurring due to fog within the scene. This study utilizes YOLOX as the base model, improving the performance of traditional YOLO models applied to single-frame blurry images. It integrates with existing multi-frame detection methods like YOLOV to achieve robust neural networks capable of stable predictions in both single-frame and multi-frame scenarios. The study enhances the preprocessing methods of existing static object detection models by incorporating new globally grayscale Gaussian blurry images for training. It optimizes the loss function to meet the predictive needs of blurry images, achieving performance improvements without requiring additional time for predicting new blurry image detection models. This approach also ensures the versatility of the new network model across various scenarios and the general applicability of the new model.