The Low-light License Plate Recognition via CNN

Abstract

Abstract The low-light license plate recognition (LPR) is an important task in LPR, and the task of low-light LPR is a challenge in LPR. Compared with ordinary LPR, low-light LPR is more challenging. The first is that there are few studies on low-light LPR, and there is a lack of dedicated datasets. Besides, there are few lightweight networks dedicated to low-light LPR. The lack of lightweight private networks makes it difficult to deploy LPR methods efficiently. Based on this, this paper proposes a low-light LPR method. Specifically, we propose a dataset dedicated to low-light LPR with a sample size of 19,121. Besides, we propose a lightweight LPR method implemented using a lightweight architecture similar to VGG. We also compare with some commonly used lightweight methods, and the comparison results show that our proposed method has better performance. In addition, we also visualized some data that are difficult for the model to identify, and found several types of challenging data. This result is helpful for subsequent researchers to improve on this type of data.