Research On Semantic Segmentation Algorithm Of Controlled Items Image Based On Deep Learning

The research of this paper is based on the semantic segmentation of the images of controlled items in the actual security inspection process.At present,most of the inspection work of controlled items is through the security inspection X-ray machine to scan the package,and human eyes judge whether the package contains the controlled item.Among them,the security inspection X-ray machine has a poor detection effect on some small item in the package,or even fails to detect,there are potential safety hazards.The automatic semantic segmentation of controlled items can effectively improve the efficiency of security inspection and improve the safety index,and at the same time,it can save labor costs and protect people’s safety in public travel and public places.Semantic segmentation of an image is one of the basic tasks of computer vision.The semantic information of an image is obtained by classifying each pixel of the image with a defined label.The pixels with similar visual features are classified into the same type of label,and each pixel is divided by to achieve the effect of regional division.The work done in this paper for the task of semantic segmentation of controlled items images is described as follows:(1)The current status of semantic segmentation based on deep learning and the main models and technologies used are described.Specially collected and produced a unique controlled items image data set CLD2020,a total of 1500,used for the research on the semantic segmentation algorithm of controlled items images.The data set contains six categories: knives,guns,scissors,axes,blunt objects,and other(small control Items,suspected dangerous items,etc.).(2)Proposed the Multi-scale Deep Convolutional Networks(MDCNet)semantic segmentation model,the Encoder part mainly uses the deep convolutional neural network to obtain the low-level information of the controlled item image,and uses the pyramid pooling layer in the network to provide high-level abstract features of the target to complete the correct classification of the pixel tags of the controlled items.The Decoder mainly uses cascaded convolution and attention mechanisms to fuse different networks and achieve feature enhancement to improve the classification accuracy of the images of controlled items.The performance of FCN,Deep Lab V2,PSPNet and Deep Lab V3+ algorithm and the MDCNet algorithm proposed in this paper are compared in the VOC2012 test set.The experimental results show that the MDCNet algorithm can achieve better performance results,which is 2.11% higher than that of PSPNet Mean Intersection over Union(MIo U),and only Deep Lab V3+ MIo U the difference is 1.29%,and the semantic segmentation performance of the MDCNet algorithm on ordinary objects is equivalent to the comparison algorithm.In the test of the CLD2020 data set,the MDCNet algorithm is 1.61% higher than the(MIo U)of the Deep Lab V3 algorithm.When the MDCNet algorithm targets the CLD2020 data set,the segmentation accuracy of small controlled item has been significantly improved compared with Deep Lab V3+.(3)The full convolutional network Seg Net,FC-Dense Net and Deep Labv3+ are studied and compared with the CLD2020 data set,and the performance of the algorithm is evaluated from the classification accuracy,actual effect and computational efficiency.Using Conditional Random Field(CRF)as a post-processing module for improved segmentation,the Pixel Accuracy(PA)of Seg Net and FC-Dense Net were increased by 4.3% and 1.1%,and MIo U was increased by 6.4% and 1.5%,respectively.In Deep Labv3+,Xception and Mobile Net V2 were used.PA increased by 10% and 0.9%,and MIo U increased by 9.4% and 0.3%.It can be seen that the performance improvement of CRF for Deep Labv3+(Xception)is the most obvious.While the network performance of the fully convolutional network is improved after the CRF is added,the computational efficiency is reduced to a certain extent.