针对传统语义分割算法参数量大、运行慢,不利于违禁品识别技术实际应用的问题,提出一种基于轻量化分割网络的违禁品识别算法。在模型的浅层特征层设计空洞卷积模块来扩大网络的感受野,减少误分类并提升分割精细度。在深层特征层设计非对称卷积模块取代传统单一串联卷积操作,降低计算复杂度。实验结果表明,所提算法在识别精度和速度上取得了均衡的性能,平均交并比(mIoU)达73.18×10 -2,每秒传输帧数(FPS)达27.1。

Aimed at the problem of traditional semantic segmentation algorithms having large parameters and slow running time, which are not conducive to their practical application for contraband identification technology, this paper proposes a prohibited item identification algorithm based on a lightweight segmentation network. A dilated convolution module is used in a shallow feature layer of the model to enlarge the receptive field of the network, reduce misclassification, and improve segmentation precision. To reduce computational complexity, an asymmetric convolution module is used in a deep feature layer to replace the traditional single convolution operation. The experimental results show that the proposed algorithm achieves balanced performance for identification accuracy and speed, the mean intersection over union (mIoU) is 73.18×10 ^-2, and the frames per second rate (FPS) is 27.1.