Research On Knife Tool Detection Method Of Airport Security Inspection Based On Statistical Learning

Although the artificial intelligence technology with deep learning has been developed rapidly in recent years,especially be existed in various applications emerge endlessly.However,at present,the detection of prohibited things is still largely dependent on manual detection.The reason is that the detection of prohibited items in public places is related to the safety of people’s lives and property,and has a heavy responsibility,so there can be no mistakes.In addition,because the placement angle of the articles in the passenger’s package and the related superposition and occlusion between the items,the existing detection methods based on computer vision are difficult to achieve the level of manual inspection,and the automatic detection methods need to be further improved.In view of the above problems,this paper proposes an airport security knife tool detection method based on convolution network feature extraction and statistical learning,it includes the following aspects:Firstly,by analyzing the characteristics of the objects in the X-ray security inspection image and the classification errors in image classification,it is considered that the classification accuracy of security inspection X-ray images depends largely on the description ability of features for object parts.In this paper,a fine-grained image classification method based on strongly supervised object components detection is proposed.By employing pre-trained convolutional nueral networks,we firstly localize the main components of the object in X-ray images,the positions of these regions are mapped to the positions of the convolution feature map,then the useful convolution descriptors are gathered together and the statistical features are extracted.Experiments confirm the effectiveness of the method for fine-grained x-ray security images classification.Secondly,presented a semi-supervised part-based fine-grained categorization method,we use saliency detection to discover set of parts and adapt local feature selection and aggregation which benefits the discrimination of the final features.We accumulate the convolution feature map to obtain the accumulation map,and cluster all the elements according to their values,dividing accumulation map into some parts.For the convolution descriptor of each local region,average down sampling is used to obtain mean vector as the local feature.Then,the correlation between local features and image features of each category is calculated,the largest of these correlation values is the weight of this region.These values are used to weight the elements in this region on the cumulative map,finally,the convolution descriptor is selected according to the weighted cumulative map to generate image features.This method can obtain the component information of the object without labeling the object components,which can be used to improve the classification accuracy,and has a fast speed.Finally,an object location method based on convolution descriptor principal component projection and a target feature extraction method based on local region convolution descriptor are proposed.The method is evaluated the correlation of descriptors and then by setting a set of thresholds obtains some category-consistent regions.Features are extracted from these regions to complete object detection.The proposed method is helpful to solve the problem of weak supervision object detection in security inspection.