Abstract: In response to the problem of low accuracy in license plate recognition due to the limitations of camera resolution and motion blur in drone-captured images, a license plate feature reconstruction and segmentation algorithm called Zoomin-Net is proposed. The algorithm consists of an adversarial perceptual image reconstruction subnet (APIR) that reconstructs degraded high-frequency details in drone-captured images to enhance image resolution. Additionally, a cross-layer encoding-decoding feature fusion subnet (CEDF) is designed to parallelly fuse shallow texture features extracted by the encoder and deep semantic features restored by the decoder using skip connections. Furthermore, a feature enhancement guiding module (FEGM) is incorporated into the backbone network to strengthen the extraction capability of license plate features and reduce network complexity through residual dense connections. A focus perceptual module (FPM) is also designed and applied to optimize the reconstruction of license plate features. Experimental results on the degraded CRPD public dataset demonstrate that the ZoominNet model achieves a 15.67 percentage points improvement in recognition accuracy compared to the YOLOv8m model, with only 9.8% of the parameter count of the YOLOv8s model. This research achievement holds significant value for the practical application of low-altitude license plate recognition using drones.

Key words: UAV image, license plate segmentation, information fusion, feature reconstruction, license plate recognition

摘要： 针对无人机视角下车牌图像受限于摄像头分辨率和运动模糊而导致车牌识别准确率低的问题，提出了一种车牌特征重建与分割算法ZoominNet。设计对抗感知图像重建子网（APIR)，重建无人机捕获图像中退化的高频细节信息，增大图像分辨率；构建与APIR并行的跨层编解码特征融合子网（CEDF)，使用跳跃连接的编码器-解码器结构，将编码器提取的浅层纹理特征与解码器恢复的深层语义特征相融合；在骨干网络设置了针对车牌特征的特征强化引导模块（FEGM)，采用残差密集连接机制提升网络对车牌特征的提取能力并缩减网络规模；设计和应用聚焦感知模块（FPM)优化车牌特征重建效果。实验结果表明，在退化的CRPD公开数据集上，ZoominNet模型较YOLOv8m模型在识别准确率指标上提高了15.67个百分点，参数量仅为YOLOv8s模型的9.8%。这一研究成果对于无人机低空车牌识别的应用落地具有重要推进价值。

关键词: 无人机图像, 车牌分割, 信息融合, 特征重建, 车牌识别