## In the past three decades various image

First, to determine the optimal parameters for Gaussian per-filter, three different kernel sizes varying from 3×3 pixels to 7×7 pixels with an increment of 2 pixels are used to smooth the third image set. Fig. 6 depicts the mean bias error and standard deviation of the displacement results calculated by Q4-DIC and Q8-DIC. It can be summarized from Fig. 6 that LDE225 per-filtered images produce smaller bias error than unfiltered ones, and larger kernel size yields smaller bias error. Specifically, Gaussian filters with a size of 5×5 and 7×7 pixels yield comparable bias errors, which are smaller than the results calculated by the 3×3 Gaussian filter. On the other hand, Fig. 6 indicates Gondwana a larger kernel size leads to larger SD errors. Thus, smaller kernel sizes (e.g. 3×3 pixels, 5×5 pixels), which keep the increase of the SD error under 10%, are preferable. On balance, the Gaussian per-filter with a size of 5×5 pixels is optimal for reducing the bias error in FE-based DIC. This conclusion is consistent with our previous work [30].