Fast Target Detection And Recognition Based On Infrared Array Scanning Equipment

With the development of machine vision,the demand for target detection and recognition technology for infrared images is also rising.Infrared array scanning system has the characteristics of strong detection ability and high scanning efficiency.Therefore,researching algorithms based on this type of equipment has become one of the hotspots in the field of infrared machine vision.Usually,the equipment outputs need to be stitching on horizontal direction for supervisor's and algorithms' convenience.However,the infrared area scan equipment could be affected by unstable scanning rate and platform vibration when working,and could not generate stable panoramic images.Besides,array scanning equipment,with high output rate and large amount of data,requires algorithm to run fast enough.What's more,due to the features of infrared image and the limitation of lame manufacturing process,the images collected by the infrared array scanning equipment have several defects such as strong noise,low signalto-noise ratio,and intensity inhomogeneity,which pose greater challenges to related algorithm.In this paper,based on the fast target detection and recognition technology of infrared array scanning equipment,three aspects are studied,including image stitching,moving target detection,and target recognition.The main contents are as follows:1.After analyzing the panoramic images generated by simple splicing of infrared area array images,the vignetting and irregular offset in area array scanned images are founded,and a set of solutions combining image registration and image fusion is proposed to solve such problem.First,this paper analyzes the existing image registration algorithms based on gray domain,transform domain,and features.Finally,the phase correlation method is used for registration,and the edge detection step is added on the beginning step,which effectively reduces the intensity inhomogeneity's impact.Then the image fusion stage selects the fade in and fade out method to implement,which effectively eliminates the stitching seams caused by intensity inhomogeneity.Experiments show that when the proposed stitching algorithm uses normalized cross-correlation as the evaluation criteria,it has higher accuracy than other algorithms,and the processing speed reaches 3.3ms/frame.2.Aiming at the shortcomings of the existing moving target detection algorithms that lack universality and the ability dealing with complex scenes are weak,a target detection algorithm based on histogram statistics is proposed.Through the analysis of common target detection algorithms such as the frame difference method and background modeling method,it is found that the existing algorithm solves the problem in a relatively narrow scope: For example,Gaussian background modeling to extract moving targets in a simple background has good performance,but it shows not the same performance under the complex background.On the contrary,KDE(Kernel density estimation,kernel density estimation method)is just the opposite.So an algorithm that can select different detection strategies according to the background type is designed.Firstly,a histogram is established for each pixel,and a stable pure background is generated by the histogram;then the noise variance of the image is obtained by threeframe difference,and the background is divided into simple background and complex background according to the noise variance and histogram;Finally,according to the background type,a single Gaussian model and an improved two-level threshold KDE are used to extract an entire foreground,which ensures the integrity of the extracted target in a complex scene.Experiments show that the proposed algorithm has the highest recall-accuracy rate up to 78.2%,and the processing speed reaches 2.9ms/frame.3.In order to address the problem that the existing classification networks based on deep learning classify infrared targets using CPU is too inefficient,a lightweight neural network is designed.Based on the structure and design skills of the existing classification network and considering infrared targets has less texture details and deep semantics,proposed network reduces the input size and convolution kernel size,and it also reduces the depth and width of the network.It greatly improves the recognition speed without losing the recognition accuracy.Experiments show that the average recognition accuracy of the proposed algorithm is 96%,and the processing speed is1.2ms/frame.