| Target detection is one of the core topics in the field of computer vision.Its purpose is to get all interested targets from the provided images.It is a prerequisite to solve some other problems,such as semantic segmentation,target tracking and so on.Compared with the traditional detection algorithms,the target detection algorithm based on deep learning shows better performance and broader development prospects.It is widely used in national defense,transportation,medical treatment and so on.In order to improve the performance of target detection algorithm in the case of insufficient visible light imaging conditions,the introduction of infrared image has become an idea to solve the above problems.Infrared image has the characteristics of long imaging distance,strong antiinterference ability and long working time.At the same time,infrared image also has the disadvantages of low resolution,more noise,blurred image and small target in the image.In order to improve the performance of infrared image in target detection algorithm based on convolutional neural network,aiming at the shortcomings of infrared image,YOLOIRS3 algorithm is proposed in this paper.The specific contents are as follows:(1)Aiming at the characteristics of weak target and more noise in infrared image,an image processing method of multi response fusion local contrast is proposed.Firstly,Gaussian Laplacian(Lo G)filter and negative Lo G filter are used to process the bright and dark targets in the image to obtain the corresponding image,then the two images are fused to obtain the enhanced bright and dark targets at the same time,and then the local contrast method is used to further enhance the data and suppress the background of the image.Compared with other baseline methods,this method plays a more obvious role in data enhancement and background suppression,and has relatively high detection rate and low false alarm rate.(2)Aiming at the characteristics of small target and lack of detail information and texture information in infrared image,a multi-scale feature fusion attention mechanism method is proposed.In addition,an efficient channel attention module is added in the feature fusion stage to improve the correlation of different channel feature maps.In this paper,experiments are carried out on self-made data sets.The experimental results show that the four layer prediction network structure and efficient channel attention module proposed in this paper improve the detection performance at the least cost.(3)According to the characteristics of infrared image target detection,some other improvements are made.Based on the real-time characteristics of target detection,the lightweight Mobile Netv2 is used to replace the backbone network of YOLOv3 to reduce the delay.According to the characteristics of infrared image gray image,it is modified into single channel data input.The infrared small target data set is made and labeled.According to the characteristics of the self-made infrared small target data set,the method of not loading the pre training model is used for end-to-end training.Finally,the modified lightweight model has fewer parameters and faster calculation speed,which basically achieves the detection accuracy of YOLOv3. |
