In-Memory-Computing (IMC) Technique in Local Difference Decision Block of an On-Board Satellite Hyperspectral Data Compression Algorithm

Due to high data acquisition rates and large data volumes, on-board hyperspectral imagery (HSI) demands data compression before transmission. The major bottleneck in on-board implementation of algorithms is the limited availability of power and resources along with the need to be less error prone while accessing and processing the data. The recently proposed simple lossless algorithm (SLA) for on-board satellite hyperspectral data compression, performs ~80% calculations by utilizing previous neighborhood pixels, which in turn requires majority of memory access. This may lead to an increase in overall power dissipation and probability of error due to bit flips in the von-Neumann architectures of SLA. Thus, to reduce the probability of error and to improve power efficiency, this work adopts in-memory-computing (IMC) technique for the local difference (LD) decision block of the SLA for non-overlapping case. The LD decision block calculates absolute difference between the neighboring pixels and requires XOR operation for the same. To perform IMC, the SRAM cell from literature is modified to get XOR between two pixels. A non-overlapping (6*16) bit memory cell is implemented using von-Neumann and IMC based architecture for LD decision block. The proposed architecture allows reduced memory overheads and instead of using more number of transistors in a cell, consumes comparable power when compared to the von-Neumann architecture. The error resilience of SLA is expected to improve, as the total memory access for processing a pixel is reduced by 21.42%.