Quantisierung eines Autoencoders zur Kompression von Satellitenbildern auf einem FPGA

Kurzfassung

The ScOSA project of DLR will develop a distributed high performance on-board computer for use on a CubSat. On this computer, neural compression methods are to be used for the compression of satellite images. In this context, the development of a prototype will be addressed, in which an autoencoder will be implemented on a FPGA. This bachelor thesis deals with the adaptation of the autoencoder in order to enable efficient executability on the FPGA hardware. The floating point operations and nonlinear functions of the compression model pose a particular challenge, as they cannot be efficiently computed by the FPGA hardware. To overcome these challenges, a complete quantization of the encoder and decoder network was performed. Since quantization leads to a loss of reconstruction performance, mission requirements were defined that specify the maximum allowed quality loss. To implement the quantization, a combination of the following approaches was required. On the one hand, PLAs were created for the non-linear functions and converted into integer calculations using fixed-point quantization. On the other hand, PTQ was applied for the quantization of the convolutional and deconvolutional operations. By combining the fixed-point quantized, linear approximations and the quantized convolutional and deconvolutional operations, a complete quantization of the encoder and decoder networks was achieved, enabling efficient execution on the FPGA hardware. Additionally, a significant portion of the mission requirements regarding the maximum loss in reconstruction quality could already be met without further optimization. To improve the reconstruction quality, a promising optimization approach was introduced, which, when implemented, could ensure compliance with all mission requirements. The quantization performed provides a solid foundation that can be built upon within the scope of the ScOSA project to enable the implementation of a compression model on the onboard computer.