针对监控视频中车牌图像超分辨率重建数据处理量较大、性能差的问题,提出了一种基于逐级反投影网络的车牌图像超分辨率重建算法。首先,为了减少超分辨率网络数据处理量,先检测并提取出低分辨率车牌区域。然后,分解深度反投影网络(DBPN)中较大的采样倍数,以逐级采样的方式完成迭代反投影。在逐级反投影单元中,跳线连接融合逐级采样产生的中间尺度特征,以提高特征利用率;用1×1卷积层降低融合后中间尺度的特征维度,同时保留关键信息。最后,根据逐级上投影单元产生的特征图重建高分辨率车牌图像。实验结果表明,相比DBPN,本算法不仅降低了超分辨率网络的数据处理量和参数量,且重建的车牌图像质量在主观感受和客观评价指标方面都得到了很大的提升。

To address the issue that large amount of data processing and poor performance of super-resolution reconstruction of license plate images in surveillance video, in this paper, a super-resolution construction algorithm of license plate image based on gradual back-projection network is proposed. First, in order to reduce the amount of super-resolution network data processing, the low-resolution license plate area is detected and extracted. Second, the larger sampling multiples of deep back-projection network (DBPN) are decomposed and the iterative back projection is completed by sampling step by step. In the gradual back-projection unit, skip connection fuses the middle-scale features produced by the gradual sampling operations to improve the feature utilization rate. 1×1 convolutional layer is used for reducing dimensions of fused middle-scale features, while preserving key information. Finally, the high-resolution license plate image is reconstructed according to the feature image generated by the gradual projection unit. Experimental results show that the proposed algorithm not only reduces the amount of data processing and parameters of the super-resolution network, but also greatly improves the subjective feeling and objective evaluation index of the reconstructed license plate image quality compared with DBPN.