Research On Controlled Tool Detection Method Based On Deep Convolutional Neural Networ

Object detection plays an irreplaceable role in security field and provides a powerful guarantee to travel.However,the lack of relevant datasets limits the research.Moreover,it is difficult for generic object detection algorithms to detect effectively due to the differences in detection devices and the interference of various factors.Therefore,it is important to study the detection algorithms and build a real-time and high-precision detection algorithm.This paper studies the knife detection in natural images.A real-time and high-precision detection scheme is proposed based on the construction of a knife detection dataset.In view of the problem that the existing methods are difficult to detect targets under the interference of a variety of factors,a highprecision detection scheme is introduced.The main work of this article includes:In order to meet the requirements of detecting knives in real scenes,a diverse knife dataset(DKD)is built.3790 knife images in the dataset are collected through Internet or surveillance videos.The position and category of the target in every image were manually labelled.The dataset contains training as well as test set,where the test set is divided into three subsets according to the detection difficulty.The constructed dataset amplifies the size and diversity of the current knife detection dataset.To solve the problem that the existing algorithms are difficult to detect knives in natural images in real time and high precision,a knife detection network scheme(MANet)combining spatial and channel attention mechanism is proposed.The residual network is used as the backbone network to improve feature extraction ability.For shallow features,a module which stacks spatial and channel attention in series is introduced to enhance the informative channels and important regions.For deep features,a kind of channel attention is introduced to improve the dependence between feature channels.Experiments conducted on the proposed dataset verify that the MANet has fast detection speed and high detection accuracy.Aiming at the problem that current algorithms are difficult to accurately detect knives under the interference of various environmental factors,the article advises a knife detection network scheme(ARFNet)based on adaptive receptive field block.An adaptive receptive field block is constructed to enhance the ability of the network to extract discriminant features.For the problem of different scales,feature pyramid structure is used to achieve multi-scale feature fusion.To suppress the inconsistency across different feature scales,a dense feature fusion block is developed to improve the scale-invariance of features.Experiments on DKD and PASCAL VOC datasets show that ARFNet achieves satisfactory detection results on both class-specific and generic objects,compared with the state-of-the-art methods.