TANet: A Tiny Plankton Classification Network for Mobile Devices

Figure 2

(a) Fire module defined in SqueezeNet. 1 × 1 and 3 × 3 represent the convolutional kernel size. 16 and 64 represent the number of filters. (b) Multibranch fire module. While keeping the number of channels unchanged, we split one branch in (a) into eight branches. (c) Group fire module. Group convolution operation has multibranch effect as (b). 8 indicates the number of convolution groups, corresponding to F in formulas (2) and (3).

(a)

(b)

(c)