基于语义分割的多目标违禁品识别算法

基于深度学习的语义分割算法可以实现安检违禁品自动识别，并获得违禁品的位置、类别及形状信息。但传统的语义分割算法在面对违禁品尺寸不一且目标多样的识别任务时表现较差。针对该问题，本文提出了一种基于语义分割技术的多目标违禁品识别算法。编码阶段，设计使用空洞空间金字塔卷积模块（Atrous Spatial Pyramid Convolution Block, ASPC），提升网络对于特征图多尺度信息的挖掘能力。同时引入注意力机制，对ASPC模块的特征提取过程进行监督，进一步提升模块的特征提取能力。解码阶段，受U-Net模型启发，采用逐级上采样操作，同时加入1×1卷积实现通道降维，减少计算量，提升模型运行速度。实验结果显示，本文提出的算法在多目标违禁品识别任务中表现良好，平均交并比（mIoU）得分78.62，处理单张图片用时（Time）68ms。 

A Multi-object Prohibited Items Identification Algorithm Based on Semantic Segmentation

The semantic segmentation algorithm based on deep learning can realize the automatic identification of prohibited items in security inspection and obtain the location, category and shape information of prohibited items. However, the traditional semantic segmentation algorithm performs poorly in recognizing prohibited items of different sizes and various objects. In response to this question, this paper proposed a multi-object prohibited items recognition algorithm based on semantic segmentation technology. In the encoding stage, the atrous spatial pyramid convolution block (ASPC) was designed to improve the ability of the network to mine multi-scale information of the feature map. At the same time, attention mechanism was introduced to supervise the feature extraction process of ASPC module and further improve the feature extraction ability of the module. In the decoding stage, inspired by U-Net model, we adopt step-by-step upsampling operation and added 1 × 1 convolution to reduce the channel dimension, the computation was reduced and the running speed of the model is improved. Experimental results show that the proposed algorithm performs well in the multi-object prohibited items recognition task. The score of mean intersection over union is 78.62, and the processing time of a single image is 68 ms.