Project Figures 3

Figure 3: Connecting the pre-processed input features in the frequency domain to ResNet-50. The three input layers (the dashed gray blocks) in a vanilla ResNet-50 are removed to admit the 56⇥56⇥64 DCT inputs. We take 64 channels as an example. This value can vary based on the channel selection. In learning-based channel selection, all 192 channels are analyzed for their importance to accuracy, based on which only a subset (<<192 channels) is used in the static selection approach.