Automatic Annotation Approach For Prohibited Item In X-ray Image Based On Deep Learning

In recent years,with the rapid development of deep learning,its application in the field of smart security inspection is triggering the upsurge of research and development of new technologies,such as prohibited item automatic detection,prohibited item instance segmentation and Threat Image Projection(TIP).At the meantime,it is universally acknowledged that deep learning approaches are based on massive data.But unfortunately,very few X-ray datasets have been published for research purposes,and manually annotating datasets is an extremely time-consuming task.For these reasons,this thesis proposes two automatic image annotation models based on two different mechanisms,aiming to minimizing the annotation time and yield standard ground truth.Most current methods treat pixel-wise annotation as a pixel-labeling problem,resulting it difficult to modify the annotation results.To solve this problem,we here cast it as a polygon prediction task,mimicking how most current datasets have been annotated,and yield the annotation by automatically producing a series of continuous vertices of the polygon outlining the prohibited item.Theoretically,the low-level features extracted during the down-sampling process are useful for generating sharp,detailed boundaries of a polygon,while high-level features are useful for grasping the overall shape and position of a polygon.Therefore,this thesis first proposes an automatic image annotation model based on adaptive multi-scale attention mechanism,named PANet model.this model introduces a feedback structure to automatically combine the multi-level features needed at the next time step according to the vertex position predicted at the current time.Although PANet model obtains the accuracy of 71.68%in automatic mode,the addition of the feedback structure made the model need to recombine features at each time step,resulting in running time long.To solve the problems of the PANet model,this thesis proposes another automatic image annotation model based on multi-path refinement mechanism,named PRN model.At each time step,the model predicts the vertex positions at different resolutions according to the different level features respectively,and then uses the multi-path refinement process to gradually refine the vertex details of the polygon to obtain accurate high-resolution results.Meanwhile,we design a mixed loss function for PRN model to eliminating the subjectivity error on account of the ground truth itself,which makes the model have strong generalization ability.To evaluate the proposed model,we present a high-quality X-ray segmentation dataset,named PIXray.The dataset contains of 5,148 X-ray images,in which 6 classes of 9116 prohibited items have pixel-wise ground truth.The experimental results show that two automatic image annotation models proposed in this thesis can achieve efficient annotating of prohibited items,but the PRN model is superior to the PANet model in all indicators,and achieves the accuracy of 81.74%in automatic mode.It speeds up the annotation process by a factor of 4 across all classes in PIXray dataset,while achieving 93.4%agreement in IoU with original ground truth after several manual fine-tuning.This thesis further proves that the two models are also effective on other natural image datasets.