Baggage Inspection Based On Deep Convolutional Neural Network

Baggage inspection based on X-Ray image can automatically identify whether there are prohibited items in the baggages.In recent years,with the development of deep convolutional neural networks,security inspection automation has made huge progress.Due to its structural advantages,it has shown great potential in processing structured data,which has further developed the X-ray image-based baggage detection technology.Owing to the fact that prohibited items in the security inspection package are often disorganized and disorderly,the prohibited items in the package are easily overlapped and severely blocked.The existing methods only use the geometric appearance information to perform detection.However,the low-level edge cues closely related to geometric information will inevitably introduce irrelevant noise in image processing.In addition,complex background information interference is always accompanied by uncertainties and model inference error,making the occlusion problem a challenge in security inspection scenarios.In addition,due to the wide-scale distribution of prohibited items in packages,the current methods show insufficient detection accuracy for small-scale prohibited items.For security reasons,we must ensure the recall rate,therefore,detector outputs plenty of false detection results,making small-scale item detection another challenge.We propose a method based on multi-scale cross-image weakly-supervised learning method in order to improve the detection performance of detectors in occlusion scene.Specifically,we propose a scale interaction module where the features in neighboring scales are interacted one or more times to alleviate the semantic gap between features.This approach could enhance model’s perception ability to distinguish between targets and nontargets.Then,we design a cross-image weakly-supervised semantic analysis module.This module could generate pseudo-labels in a weakly supervised manner and utilize co-attention mechanism to perceive similar and different targets,breaking through information bottleneck of isolated detection of a single image.Finally,we introduce a multi-task learning module.In detail,detector could use an additional image segmentation branche to perform pixel-level segmentation task and provide semantic-level information to assist the final baggage detection.Compared with the current X-Ray image baggage inspection methods,our method has obtained m AP improvements of 2.26% and 2.49% on the SIXray and OPIXray datasets,respectively.We propose a cascaded detector combined with hard-example enhancement so as to improve the detection ability of detectors for small-scale prohibited items.Particularly,we propose a data augmentation method with hard-example mining to handle small-scale prohibited item problem.Then,we design a cascaded detector with an evaluation net,which filters out wrong results from initial detector.Compared with a complex detector,it can effectively mitigate training pipeline,stabilize recall rate,and improve detection accuracy.We achieved 2.64% and 2.07% improvements in mean Average Precision(m AP)on SIXray and OPIXray datasets,respectively.We propose a method to obtain different fine-grained model training labels to alleviate the problem of lack of labeled data.In particular,our method constructs pseudo-masks for image segmentation and image-level labels by only using original object detection labels for model training.Meanwhile,we reasonably use the training process to generate multi-stage models,and generate new training data by recombining different samples and labels.Our approach could avoid the problem caused by insufficient labeled training data.In addition,our methods may of independent interest,which makes it applicable to various deep learning-based scenarios.The prohibited object detection methods based on deep convolutional neural network we propose aim to improve the safety of public places.Through real-time and accurate object detection for security inspection packages,our model can quickly determine whether there are prohibited items in passengers’ luggage and packages.