X射线安检违禁品智能检测模型复杂场景数据

本数据集包含种类丰富、复杂多样场景下的14类违禁品的X光安检图像，通过对图像的标注、抠图、增强、融合等处理，可作为优质样本训练生成智能检测X射线安检违禁品的算法模型，实现对藏匿在其他物品中或伪装成其他物品等复杂环境下的14类X射线安检违禁品的精准识别。1）数据来源：原始数据使用自研X光安检设备自行采集得到； 2）数据清洗和标注：对收集到的数据使用传统 图像处理 方法，包括灰度化、二值化、腐蚀膨胀等操作去除包含噪声的图片数据；利用半自动标注工具得到伪标签，然后使用人工修正标注，并设置审核机制，保证标注的准确性和一致性。 3）违禁品数据集构建：构建一个包含烟花爆竹的违禁品数据集，用于在训练和推理过程中学习罐装易燃气体，气雾剂，压缩装气体，瓶装液体、打火机、电击器、管制刀具、锂电池、枪支及仿制品、日常刀具、手铐、蓄电池、烟花爆竹、子弹及仿制品等14类品项的特征，包括边缘、形状、颜色等。 4） 深度学习 架构选择：选择适合的视觉检测模型，本算法基于FasterRCNN模型构建。 5）模型训练：在标注好的数据集上训练模型，通过监督学习的方式让模型学习识别罐装易燃气体，气雾剂，压缩装气体，瓶装液体、打火机、电击器、管制刀具、锂电池、枪支及仿制品、日常刀具、手铐、蓄电池、烟花爆竹、子弹及仿制品等14类品项的特征。使用交叉验证和不同性能指标（如准确率、召回率）评估模型的识别能力。 6）超参数调优：进行超参数调优，包括学习率、模型结构和尺寸、目标损失函数等，以优化模型性能。 7）模型验证：在独立的测试集上验证模型的性能，确保模型在未见数据上也能表现良好。 8） 智能检测结果生成：通过训练好的模型和违禁品数据集，通过循证规则来完成上述14类违禁品的智能识别，最终基于算法模型输出相关属性，具体包括品项类别、品项位置和置信度；输出的置信度与阈值进行比较，如果置信度大于阈值，且品项类别为上述违禁品的具体类别，则表示检测出违禁物。示例性地，样例数据的X光安检图像中，检测出3种类别的违禁品，目标物数量记为3，品项类别分别为压缩装气体、瓶装液体和锂电池，品项位置分别为[661, 210, 750, 335]、[373, 323, 432, 422]、 [383, 162, 494, 292]和[744, 539, 769, 551]。

This dataset contains X-ray security inspection images of 14 types of contraband in various and complex scenarios. Through processing such as image annotation, matting, enhancement and fusion, it can serve as high-quality samples to train intelligent X-ray security contraband detection algorithm models, enabling accurate identification of 14 types of X-ray security contraband hidden in other items or disguised as other items in complex environments.
1) Data Source: The original data was collected using self-developed X-ray security inspection equipment.
2) Data Cleaning and Annotation: Traditional image processing methods including grayscale conversion, binarization, erosion and dilation operations were applied to remove image data containing noise. Pseudo labels were obtained via semi-automatic annotation tools, followed by manual annotation correction and the establishment of an audit mechanism to ensure the accuracy and consistency of annotations.
3) Contraband Dataset Construction: A contraband dataset including fireworks was constructed to learn the features (including edges, shapes, colors, etc.) of 14 categories of items during training and inference, namely canned flammable gases, aerosols, compressed gases, bottled liquids, lighters, electric shock devices, controlled knives, lithium batteries, firearms and their replicas, daily knives, handcuffs, storage batteries, fireworks, ammunition and their replicas.
4) Deep Learning Architecture Selection: A suitable visual detection model was selected, and this algorithm was built based on the Faster RCNN model.
5) Model Training: The model was trained on the annotated dataset. Through supervised learning, the model was trained to learn the features of the above 14 categories of items. Cross-validation and different performance indicators (such as accuracy, recall rate) were used to evaluate the model's recognition ability.
6) Hyperparameter Tuning: Hyperparameter tuning was conducted, including learning rate, model structure and size, target loss function, etc., to optimize model performance.
7) Model Validation: The model's performance was verified on an independent test set to ensure that the model performs well on unseen data.
8) Intelligent Detection Result Generation: Using the trained model and the contraband dataset, the intelligent recognition of the above 14 types of contraband was completed through evidence-based rules. Finally, relevant attributes were output based on the algorithm model, specifically including item category, item position and confidence. The output confidence is compared with a threshold. If the confidence is greater than the threshold and the item category is a specific category of the aforementioned contraband, it indicates that contraband has been detected. For example, in the X-ray security inspection image of the sample data, 3 types of contraband were detected, with the number of target items recorded as 3. The item categories are compressed gas, bottled liquid and lithium battery respectively, and the item positions are [661, 210, 750, 335], [373, 323, 432, 422], [383, 162, 494, 292] and [744, 539, 769, 551] respectively.