Research And Implemention Of Anchor-free Object Detection Based On Deep Learning

Object detection is one of the basic tasks in the field of computer vision,and it is also the basis of visual tasks such as instance segmentation and object tracking.In recent years,object detection has gradually changed from object detection algorithm based on artificial design features to object detection algorithm based on deep learning.Due to the imbalance of positive and negative samples and a large number of redundant frames in the object detection algorithm based on anchor,the object detection algorithm based on anchor-free is gradually emerging,such as Corner Net,Center Net and so on.This paper comes from a national key research and development program.Aiming at the problems of bad weather conditions,complex environment and limited visible light imaging,infrared object imaging technology is used to obtain object detection images to adapt to allweather geographical environment.However,the infrared image has poor resolution and low contrast,which brings great challenges to target detection.In this paper,based on Center Net network model,aiming at the problems of infrared image occlusion and small target detection,a new infrared anchor-free object detection algorithm is proposed.(1)Aiming at the problem of poor effect of the small object in infrared image,this paper proposes a multi-scale infrared object detection algorithm based on attention with multiscales Center Net.The infrared image has the character of low contrast and poor visual effect,so the algorithm is difficult to identify infrared object effectively.By introducing attention,the feature extraction ability of the network for the infrared image is improved.Multi-scale module is introduced to improve the small object detection effect from different scale feature maps.Combined with the attention and multi-scale fusion module,a reasonable network structure is designed to improve the detection effect of small object and the robustness of the model to the change of object size.(2)Aiming at the problem of infrared image occlusion and small object detection,this paper proposes an algorithm of attention with multi-scales time series Center Net.The current detection algorithm is to extract a single frame,detect and classify the operation,does not take into account the timing information and target motion state,ignoring the detection results of historical frames.By combining the motion model with the Io U adaptive module,we make full use of the objects motion information and historical frame detection results,improve the algorithm process,and enhance the detection effect of occlusion and small objects.(3)Aiming at the problem that the network model is limited in the actual environment,this paper proposes an improved algorithm based on the knowledge distillation.Limited by the actual environment,some deep complex network models can not be used.Through the CMC-KD algorithm,the large-scale complex network is taken as the teacher model,and the small-scale simple network is taken as the student model.The performance of the teacher model is transferred to the student model,so that the student model has the performance of the teacher model.On the premise of ensuring the performance,it reduces the requirements of the network model on environment and improves the practicability of the algorithm.In order to verify the effectiveness of the method,four algorithms,Faster-RCNN,Center Net,YOLOv3 and Corner Net,are tested on FLIR-PC dataset.In terms of detection accuracy and small target detection,AMST-Center Net algorithm performs best.Compared with Center Net algorithm,AMST-Center Net improves the detection effect of the infrared image by about3.5% and the small object detection effect by about 2.6%.The results show that AMSTCenter Net algorithm can improve the occlusion detection accuracy by 2.9% compared with Center Net algorithm.Finally,CMC-KD algorithm is used to improve the practicability of the network model while ensuring performance through knowledge distillation.