This paper proposes a deep learning-based Optical Character Recognition (OCR) system aimed at addressing the limitations of traditional Vehicle Identification Number (VIN) recognition methods in complex environments. As the unique identifier for vehicles, the VIN plays a critical role in vehicle management, registration, and tracking. However, conventional recognition methods, which rely on manual transcription or simple license plate recognition, are inefficient and prone to errors, especially in situations where there is insufficient lighting, reflections, dirt, or significant tilt angles. To overcome these challenges, the proposed system uses high-resolution cameras to capture vehicle images and applies image preprocessing techniques such as grayscale conversion and binarization to enhance image quality, ensuring that characters are clearly visible even in challenging conditions.At the core of the system is the integration of deep learning models, including Long Short-Term Memory (LSTM) networks, which automatically learn and extract key features from the images, enabling precise VIN recognition without the need for manual intervention. Compared to traditional methods based on template matching or rules, deep learning models offer greater generalization capabilities, allowing for high-accuracy recognition under various complex conditions. Additionally, the system incorporates a character verification function to ensure that the recognized VIN conforms to standard formatting and effectively distinguishes between easily confused characters, such as "0" and "O." This feature not only improves recognition accuracy but also helps prevent the misuse of charging cards, further optimizing the management and utilization of corporate vehicle resources.