Research On Long Short Term Memory Network-based Remote Sensing Image Compression Method And Hardware Acceleration

Remote sensing images are widely used in many fields such as agriculture,industry,military,commercial,etc.Howerver,with the development of information science and technology,there is a dilemma between a large amount of remote sensing data and the limited storage /transmission resource.Image compression technology is an effective method to solve this problem.Remote sensing image presents the characteristics of complex local texture and low correlation,increasing the difficulty of remote sensing image compression.This paper focuses on the problems of remote sensing image compression on embedded platform in terms of high compression ratio,high fidelity,limited resources and real-time.This paper proposes an optimized lossy compression network model for remote sensing image and its hardware acceleration implementation.On the level of remote sensing image compression algorithm,this paper analyzes the characteristics of remote sensing images and the requirements of remote sensing image compression tasks.This paper studies the compression algorithm based on the long short-term memory network（LSTM）image lossy compression method with the capability of reidula error iterative encoding,and analyzes the limitations of the original algorithm from the perspective of algorithm principle and application scenarios,and finally proposes a patch-wise fidelity controllable lossy compression method（PFCLC）.The experimental results show that this algorithm has better visual quality than the traditional algorithms at high compression ratio,and appears of error resilience,controllable fidelity.On the level of algorithm implementation optimization,this paper analyzes the necessity and feasibility of PFCLC network model optimization on embedded platform,and adopts compact network design and model quantization technology to optimize the PFCLC network.In compact network design,a lightweight PFCLC network structure is proposed based on depthwise separate convolution.In model quantization design,a convolution quantized LSTM model is proposed,and a joint positive and negative weight-feature map quantization method is designed to reduce the quantization loss.Finally,this paper integrates these two optimization methods to design a convolution quantized lightweight PFCLC network.The results show that under the same compression ratio,compared with the original algorithm,the amount of computation is reduced by95.29%,and the amount of data is reduced by 92.56%,with negligible reduced visual quality.On the level of algorithm hardware acceleration,according to the algorithm structure,this paper proposes a SOC FPGA-based PFCLC network acceleration architecture,which consists of computing reasoning module,data storage module and network control module.Moreover,this paper proposes a hardware resource-aware network parallel architecture search algorithm for searching the network architecture parallelism under limited resources.To verify the effectiveness of this architecture,this paper deploys the PFCLC network on Xilinx ZU9 EG platform.The experimental results show that,under the premise of effective compression performance,for 128*128 resolution images patches,the processing frame rate can reach up to 120 fps with power consumption of only 6.4W.Therefore this encoder has the ability to be applied to remote sensing image compression tasks on embedded platform.