Fig. 2.

The schematic illustration of the designed sASPP. (a) sASPP consisted of four atrous convolutions with dilation rates ranging from 6 to 24. The kernel size was 1, 3, 3, and 3 respectively. (b) sASPP in a certain state at the stage of training. Dilation rate and kernel size were the same as (a). $b_i$ denoted whether the feature map existed, where $b_i=0$ was defined as discard state and vice versa. (c) sASPP at the stage of testing. All the feature maps generated by four atrous convolutions were reserved at the stage of testing in terms of their magnitudes scaled by the retainable probability during the training stage.