研究目的
To implement a convolutional neural network (CNN) based hyperspectral classification on an FPGA platform for onboard processing, addressing the limitations of GPU platforms in terms of space radiation and power supply issues.
研究成果
The FPGA-based implementation of CNN for hyperspectral classification is validated through simulation results over the Pavia University dataset, showing coincidence with GPU-based implementation. The designed hardware model efficiently utilizes FPGA's flexible architecture, fast computing speed, and low power consumption, making it suitable for onboard processing.
研究不足
The implementation focuses only on the forward classification step of CNN, with the training step conducted off-line on a GPU platform. The calculation precision on FPGA, while high, may slightly differ from GPU due to the conversion of trained parameters to 32-bit single-precision floating-point numbers.
1：Experimental Design and Method Selection:
A hardware model for CNN is designed using hardware description language, focusing on the forward classification step. This includes the computation structure for CNN, implementation of different layers (convolutional layer, batch normalization layer, dense layer), weight loading scheme, and data interface.
2：Sample Selection and Data Sources:
The Pavia University dataset, acquired by the ROSIS sensor, is used for evaluation. It contains 610×340 pixels with a ground resolution of 1.3 meters and 103 spectral bands.
3：3 meters and 103 spectral bands.
List of Experimental Equipment and Materials:
3. List of Experimental Equipment and Materials: FPGA platform is used for implementing the CNN model. The model involves thousands of parameters stored on on-chip ROM and uses FIFO for buffering intermediate data between modules.
4：Experimental Procedures and Operational Workflow:
The CNN model is implemented on FPGA with a pipelined structure to maximize computational efficiency. The model's performance is evaluated by comparing its classification results with those obtained from a GPU platform.
5：Data Analysis Methods:
The classification accuracy is measured in terms of class-specific accuracy, overall accuracy, and average accuracy. The calculation precision and performance of the FPGA implementation are also analyzed.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
