Fig. 2.

Architecture of the EfficientNetB0 backbone used in our transfer learning model. Blue and red blocks represent mobile inverted bottleneck convolution (MBConv) blocks (see Fig. 3), with a kernel size of 3x3 and 5x5, respectively. The light gray block represents a strided convolution followed by batch norm and a swish activation, while the dark gray block represents an MBConv block with expansion factor 1 and kernel size 3. The resolution at each level is displayed in bold and the number of filters (output channels) is displayed with the ‘f’ suffix