| With the rapid development of artificial intelligence,neural network algorithms have strong self-learning ability and can be applied to object classification,detection and segmentation.Compared with the traditional image processing recognition technology,the object detection algorithm based on deep learning has shown good characteristics in many practical application scenarios.At present,the accuracy of Radar waves can reach the level of millimeter,among which Synthetic Aperture Radar(SAR)is the most widely used and has strong ground penetration.It can efficiently detect all kinds of targets with deep hiding and strong camouflage,and generate high resolution gray images similar to optical photos.In recent years,the number of high resolution images generated by SAR increased rapidly,so researchers began to study the processing of SAR images.In the ground battlefield,the difficulty of target recognition is greatly increased because of artificial camouflage,complex environment and the problem of distant target recognition.In this paper,the YOLOv5 based lightweight high-precision model is designed.By introducing attention mechanism,single target and multi-target are studied respectively,which improves the accuracy of multi-target and small target in complex background.The main work of this paper is as follows:(1)At present,the image features of small target and multi-target objects in targets detection are not sufficient and complete.By introducing the attention mechanism,YOLOv5 algorithm is improved.The experimental results show that the CBAM attention mechanism module is superior to the CA attention mechanism module,and the CBAM attention mechanism can make the extracted target features more perfect,and solve the recognition problems existing in small targets and overlapping condensed targets.(2)A lightweight convolutional neural network algorithm based on YOLOv5 is proposed to solve the problem of fast detection based on high precision in current target detection.By adopting the trunk network of deep separable convolution and model quantization,the balance of accuracy and lightness is achieved compared with the basic YOLOv5.(3)Expand the MSTAR data set with single target and simple background,and make a single target data set containing more than 30,000 samples.Multi-object images of different kinds of complex scenes were constructed and 1200 multi-object data sets of complex backgrounds were made.It solves the problem of lack of SAR image data.This paper is based on Ubuntu 20.04.3 LTS system and uses PyTorch framework to build the algorithm implementation environment. |
