基于热图感知的复杂场景港口箱号检测算法

Container Code Detection Algorithm Based on Heatmap Perception inComplex Scenes

受自然场景下光照、遮挡和多方向多尺度等情况的影响,港口场景下集装箱号码的精准检测仍然面临着巨大的挑战。为解决箱号检测中多余背景和干扰字符问题,根据集装箱号文本字符的结构规律,设计了一个基于热图感知的像素级箱号检测架构。通过融合CBAM的MobileNetV3轻量级初定位网络获取目标箱号预选区域,构建基于Transformer的像素级字符区域自适应网络,从预选区域中精准提取箱号感兴趣区域。基于实际数据集的结果表明,所提算法能够达到97.5%的箱号定位准确率,并能达到25帧/秒的实时性,满足实际的场景需求。

Affected by lighting,occlusion,multi-directional and multi-scale conditions in natural scenes,the accurate detection of container code in port still faces huge challenges.In order to solve the problem of redundant background and interfering characters in container code detection,a pixel-level container code detection framework based on heat map perception is designed according to the structural rules of container code characters.The pre-selected region of the target container code is obtained by the MobileNetV3 lightweight coarse positioning network which incorporates CBAM,and a pixel-level character region adaptive network based on Transformer is constructed to accurately extract the region of interest for the container code from the pre-selected region.The results based on the actual data set show that the proposed algorithm can achieve 97.5%container code positioning accuracy,and can achieve 25 frame/s real-time performance,which meets the actual scene requirements.