Research On Intelligent Detection Algorithm Of Infrared Small Object Based On Convolution Neural Network

With the improvement of the computing power of the deployment platform,deep learning algorithms can be more deployed in practical applications.In order to improve the embedded detection accuracy of infrared small targets,and obtain a lightweight model with easier deployment and faster speed,the precision optimization method of deep convolution network based on attention mechanism and the compression acceleration method of convolution network based on shallow network pruning were mainly involved.The transplantation and optimization scheme of model department was explored,and the algorithm effect was experimented.The exact work of this paper is as listed below:(1)In order to facilitate the training and testing of small target detection algorithm on small embedded platform,an optimization scheme of yolov5 s detection algorithm was proposed by using attention model,including preprocessing level,receptive field level and backbone network structure,and a multi-scale attention feature fusion method for shallow information transfer was given.On the established infrared small target detection data set,the detection effects of different models under different parameters were tested.Compared with the original network,the output feature map was redistributed,the influence of feature aggregation and interference was decreased,and the small target feature was strengthened.The improved model could advance 3% on m AP@0.5 and 2% on m AP@0.95.(2)On the basis of summarizing the implementation methods of YOLO tiny series network,the model compression and acceleration method based on shallow network pruning was studied.Starting from YOLOv5 s model,a method for detecting model compression and acceleration was given,which involves network structure,convolution mode,backbone network and model pruning.On the premise that the accuracy decreased by 4.3%,the volume of the lightweight model decreased by 86.6%and the speed increased by 53.1%.(3)A model transformation and an operator replacement method scheme of YOLOv5 s model deployed on an small embedded platform was proposed.Then the speed optimization scheme of convolution network model in embedded platform was explored,involving network structure and convolution mode.The speed of the final model with the input scale of 640 × 512 pixels could reach 209 FPS in the test of embedded platform.The loss of model quantization accuracy was reduced by using a quantization scheme with mixed accuracy of 8-bit and 32-bit hybrid quantization.A full picture detection scheme with windows to improve the minimum detectable scale was further explored.The model could correctly detect partially occluded targets as well as targets under low background temperature difference,and the minimum detectable scale is 7 × 4.