Research On Object Detection Based On Convolutional Neural Network Under Laser Active Imaging

In recent years,active imaging system has been continuously developed.As a recognition aid of reconnaissance system,it is widely used in the field of national defense and security.Compared with the passive imaging system,the active imaging system uses its own light source to illuminate the target area and receive the reflected signal for imaging.These systems have some unique technology and application advantages,such as long detection distance and high-resolution imaging,making them an attractive substitute or at least a supplementary component of the existing infrared imaging technology.Laser is often selected as the light source of active imaging system because of its high brightness,monochromaticity and high collimation.There are many methods to realize active imaging system,and one of the mature ways is to combine infrared imaging technology.Firstly,an infrared imager is used for wide field of view search.Then the laser is used to illuminate the suspicious area and the camera receives the reflected signal for imaging to achieve narrow field of view identification.Designing corresponding algorithms to process the images acquired in this process is an important step for the system to achieve high-precision object detection.Therefore,on the basis of investigating the research status at home and abroad and understanding the imaging characteristics of infrared images and laser active illumination images,according to the needs of different stages,this paper carries out research in three aspects:infrared wide field of view retrieval,laser active illumination narrow field of view detection and key parts detection of specific targets,so as to realize the corresponding detection algorithm.The work of this paper can be summarized as follows:1.Aiming at the wide field of view search process,a framework for joint image super-resolution and object detection tasks is proposed.Compared with visible images,infrared images can only be low-resolution gray images,which brings difficulties to object detection.In view of this,we add an auxiliary branch of image super-resolution on YOLOv3 to construct a multi-task deep learning model.The weights between tasks are adjusted adaptively through learning,which brings improvements in model efficiency and performance compared with manual design.The experimental results show that the addition of super-resolution auxiliary task branches is helpful to improve the performance of the detection model.2.A high-precision and real-time object detection algorithm is achieved for laser active illumination images.In the process of laser active imaging,the coherence of the laser will introduce speckle noise,atmospheric turbulence and the vibration of bearing platform will blur the image and reduce the imaging quality,which will degrade or even fail the performance of most traditional object detection algorithm based on gradient,color and texture features.In order to solve this problem,we introduce the popular deep convolution neural network algorithm as the basic framework to realize object detection.Furthermore,in order to solve the contradiction that the deep learning framework needs a large amount of annotation data and collecting laser active illumination images is time-consuming and laborious,we construct the simulation datasets of specific targets under different backgrounds by simulating the laser active illumination imaging process.The algorithm is trained on the simulation dataset,and then tested on the real dataset to verify the transferability of the knowledge learned from the simulation dataset,so as to complete the migration application of deep learning in the field of laser active imaging,and realize the UAV detection performance of 53.4% AP and 104.1 FPS based on YOLOv5 s.3.A lightweight model is designed to realize the key parts detection of a specific target.After detecting a specific target,further detection of the key parts of the target is an important link to achieve the ultimate blinding effect.Compared with conventional detection tasks,the key part detection task is simple,and does not need a large-scale high-performance detection model.In response to the needs of this scenario,we improve YOLOv5 from the perspective of model lightweight.On the one hand,we use the Shuffle module to replace the backbone of the original model;on the other hand,the detection scale range of the model is cut from three scales to two scales.The designed model is verified on the previously obtained target images.The experimental results show that the speed of the model can reach 153.84 FPS while maintaining 99.5%AP on the self-built dataset.The accuracy of cascaded UAV and its key parts detection is 4.8% higher than that of simultaneous detection,while the speed can still reach 62.5FPS,meeting the system requirements.