基于权重共享的轻量化注意力卷积编码器OA
Lightweight attention convolutional encoder based on weight sharing
环境的复杂多样导致车载传感器获得的图像存在大量的干扰信息,给交通标志的识别带来了较大困难.为此,提出一种基于权重共享预训练的轻量化注意力卷积编码器.首先,结合注意力机制与非对称卷积结构,对已编码的重要特征进行关注;然后,利用对比学习中的并行结构对编码器进行预训练,提升编码器在复杂环境下的样本学习能力;最后,对输出的特征编码进行预测,完成对交通标志的识别.实验结果表明,所提算法在模糊和遮挡条件下的准确率分别达到94.65%、91.23%,具有实用性.
Due to the complexity and diversity of the environment,there is a large amount of interference information in the images obtained by vehicle mounted sensors,which poses significant difficulties for the recognition of traffic signs.A lightweight attention convolutional encoder based on weight sharing pre-training is proposed.Important features of the encoded features are focused by combining attention mechanisms with asymmetric convolution structures.The parallel structure in contrastive learning is used to pre-training the encoder,so as to improve its sample learning ability in complex environments.The output feature encoding is predicted to complete the recognition of traffic signs.The experimental results show that the accuracy rates of the proposed algorithm under blurry and occluded conditions can reach 94.65%and 91.23%respectively,which is practical.
肖芷翊;奚峥皓
上海工程技术大学 电子电气工程学院,上海 201620上海工程技术大学 电子电气工程学院,上海 201620
信息技术与安全科学
通道注意力机制卷积编码器非对称卷积结构权重共享对比学习图像识别
channel attention mechanismconvolutional encoderasymmetric convolution structureweight sharingcontrastive learningimage recognition
《现代电子技术》 2026 (2)
44-48,5
国家自然科学基金项目(12104289)
评论