Performance Improvement of Cooperative Spectrum Sensing Based on Dequantization Neural Networks

When cognitive users perform cooperative spectrum sensing (CSS), they can transmit their sensing information at various resolutions, ranging from binary to full-precision, according to reporting schemes. The trade-off between the quantity of information and signaling overhead in reporting schemes can pose challenges for unlicensed cognitive users. In this letter, we propose a method to dequantize the low-bits sensing information based on convolutional neural network (CNN) to improve CSS performance without extra signaling overhead. The dequantization CNN takes low-bits information as input then, through regression, produces an output that approximates the full-precision version of the information. Additionally, our proposed network can function as a module regardless of the type of CSS networks. To verify the effectiveness of dequantization, we compared the distribution of output values with the distribution of target full-precision values using Kullback-Leibler divergence. Finally, we show that the performance of CSS can be improved by the proposed dequantization CNN.