Research And Implementation Of Infrared Image Recognition Based On Convolutional Neural Network

Infrared image recognition technology has been developed and valued nowadays because of the particularity and diversity of its application.The development and application of infrared image recognition technology based on neural network makes the technology get better application feedback in intelligent vehicle,all-weather monitoring,remote command,military investigation and so on.At the same time,Unmanned Aerial Vehicle(UAV)technology has developed fast and makes a new remote sensing platform by combining with infrared image recognition.The remote sensing platform has gradually become a hotspot.UAV moves fast and far away from the target,so the target pictures taken by remote sensing platform is small in size,and has large background noise.According to this situation,a system to recognize infrared image by method of the deep learning is designed in this paper:Firstly,the working mechanism of infrared image recognition system using traditional neural network is researched,and the limitation of traditional neural network in infrared image recognition is pointed out.Then a deep learning method scheme is proposed to solve the problem.A 6-layer convolution neural network is designed,and small batch training method is applied to obtain an optimized structure.After that hardware design architecture of convolution neural network is studied.Secondly,an improved design of excitation function of convolution neural network is proposed.General defects of excitation function are existing in convolution layer in neural network.For example,the ReLUs function is insufficient to express the model,the ReLUs family function will easily lead to the necrosis of neurons and Softplus function lacks sparse expression ability.To make up for the above deficiencies,a new nonlinear modified function named PRelus-Softplus is proposed.Relevant experiments are carried out on MNIST and CIFAR-10 and the results show that the improves excitation function improved the convergence speed and the recognition accuracy of the image recognition.The experiment on the infrared image database has also got a wanted result.Thirdly,by comparing methods of approximation,Table-driven Lookup method is proposed which is a combination of table lookup method and linear approximation method.Approximation scheme for PRelus-Softplus function and its derivatives is designed with advantages of high precision,good performance in real-time calculation and little occupation of hardware resources.At last,a hardware platform suitable for unmanned aerial vehicles(UAV)is designed as well as FPGA architecture for convolutional neural network.What's more,parallel computing ability and multistage assembly line computing ability are developed deeply this paper.Then a two-dimensional convolutional cache structure is designed,by which pixels of the input image are reused and therefore higher throughput capacity and reduced bandwidth requirements for outside storage are obtained.