The task of traffic line detection is a fundamental yet challenging problem in computer vision. Previous traffic line segmentation models either tend to increase the network depth to enhance the representation ability to achieve high accuracy, or tend to reduce the number of model layers or hyper-parameters to achieve real-time efficiency, but how to trade off high accuracy and low inference time is still challenging. In this paper, we propose a reinforced attention method (RAM) to increase the saliency of traffic lines in feature abstraction, using RAM to optimize the model can achieve better traffic line detection accuracy without increasing inference time. In the RAM processing, we define the line to context contrast weight (LCCW) to represent the traffic line saliency in the feature map, which can be calculated by the ratio of the traffic line energy to the total feature energy. After LCCW calculation, we add a RAM loss item to the total loss in backward processing, and then retrain the model to obtain the new parameter weights. To validate RAM on real-time traffic line detection models, we applied RAM to seven popular real-time models and evaluate them on two popular traffic line detection benchmarks (CULane and TuSimple). Experimental results show that RAM can increase line detection accuracy by 1–2% on the CULane and TuSimple benchmarks, and the ERFNet and CGNet almost reach state-of-the-art performance after the models are optimized by RAM. The results also show that RAM can be applied to the optimization of almost all encoder–decoder-based models, and the optimized models are more robust to occlusion and extreme lighting conditions.

交通线检测任务是计算机视觉中一个基本但具有挑战性的问题。以往的交通线分割模型要么倾向于增加网络深度以增强表示能力以实现高精度，要么倾向于减少模型层数或超参数以实现实时效率，但如何权衡高精度和低推理时间仍然具有挑战性。在本文中，我们提出了一种增强注意力方法（RAM）来增加特征抽象中交通线的显着性，使用RAM优化模型可以在不增加推理时间的情况下获得更好的交通线检测精度。在 RAM 处理中，我们定义 line to context contrast weight (LCCW) 来表示特征图中的交通线显着性，可以通过交通线路能量与总特征能量的比值来计算。LCCW计算后，我们在后向处理的总损失中加入一个RAM损失项，然后重新训练模型，得到新的参数权重。为了在实时交通线检测模型上验证 RAM，我们将 RAM 应用于七个流行的实时模型，并在两个流行的交通线检测基准（CULane 和 TuSimple）上对其进行评估。实验结果表明，RAM 可以在 CULane 和 TuSimple 基准上将线检测精度提高 1-2%，并且在模型通过 RAM 优化后，ERFNet 和 CGNet 几乎达到了最先进的性能。结果还表明，RAM 可以应用于几乎所有基于编码器-解码器的模型的优化，