Comparing the robustness of deterministic and stochastic edge detection circuits to transmission noise

This paper studies how the use of Stochastic Computing affects the accuracy of an edge detection algorithm. The study aims at comparing a deterministic and stochastic implementation of a Sobel filter. The comparison focuses on both implementations' robustness to output bit-flip errors, which can occur during data transmission. The mean absolute error (MAE) and root mean squared error (RMSE) metrics are used to analyze both solutions in a subset of 2,500 images obtained from a public dataset. Compared to the deterministic implementation, the stochastic circuit presented an average MAE = 1.41, and an average RMSE = 2.47. The analysis considers 1%, 2% and 5% rates for the output bit-flip error. Both implementations yield similar MAE results, but the stochastic solution shows an RMSE up to one order of magnitude lower than the deterministic one. These results imply that a stochastic implementation is preferable instead of a deterministic one in applications prone to output bit-flip errors. It also means that the stochastic solution may be applicable for low-power IoT devices and sensors for cases where approximate results are allowed.