GPU Accelerator Design For Hyperspectral Image Anomaly Detection

As an unsupervised deep learning algorithm,Stacked Autoencoder(SAE)has been applied in hyperspectral image(HSI)anomaly detection for its excellent nonlinear fitting ability.However,and the long time-consuming caused by large amount of calculation have become a major bottleneck restricting its application.Relying on GPU's massive parallel computing capability and high memory bandwidth,the GPU accelerator for HSI anomaly detection is designed.The paper's work can provide a feasible solution for real-time detection of abnormal targets in HSI.It can also provide reference for the design of GPU accelerators for other machine learning algorithms.The work starts from SAE training process and inference process.As for long time-consuming during the training process,the GPU server is used to carry out the design of training accelerator.As for the need of low-latency requirement and the limited power consumption of the computing platform,the embedded GPU is used to perform the design of the inference accelerator.The main works are as shown below:Firstly,the principle of SAE-based HSI anomaly detection is analysed and the SAE-based detector is designed.Two real hyperspectral image data are used to verify the performance of the model.The SAE-based detector achieves AUC values of 0.8569 and 0.9248 respectively under the two scenes,which are higher than the reference detector RXD.It verified the rationality and correctness of the SAE-based HSI anomaly detector.Sencondly,for the time-consuming during the training process of HSI anomaly detection model based on SAE,the training accelerator based on CPU+GPU heterogeneous system is designed.In order to improve the computational efficiency,three strategies that host computing and device computing overlapping,data transmission and device computing overlaping and mulity devices computing overlapping have been proposed in CPU and GPU heterogeneous system.The experimental results of computational efficiency test based on real HSI data show that the GPU training accelerator proposed in this paper achieves a 27 times acceleration compared to a single-core CPU at the fastest convergence speed of the model,validating its acceleration effect.Finally,aiming at the contradiction between the real-time requirement of HSI anomaly detection process and the volume,power consumption constraint of the computing platform,the GPU inference accelerator is designed based on embedded GPU.The computational delay is shortend by model computational graph reduction and the merging of the GPU kernel functions.Contrastive experiment results based on real HSI data for computational efficiency and power consumption show that the inference accelerator proposed in this paper can achieve 157 times the speedup ratio and 113 times the energy efficiency ratio compared to the ARM processor,and 71 times the acceleration ratio and 119 times the energy efficiency ratio compared to the eightcore DSP.