Figure 8.

(a) Traditional CNN unit mode and (b) DenseNet unit mode. (a) Input gradient information in conventional unit module goes through many layers. (b) DenseNet unit module uses dense jump connection to splice input features with learned features as subsequent layer inputs thus achieving feature reuse, which is easy to train and more parameter efficient in DenseNet. Specific structural design details can be expressed in following equation with some notations defined in advance: input picture is X₀, l is index of neural network layers, H(•) indicates nonlinear operations combination, and H(•) in DenseNet refers to composite function consisting of BN, ReLU, and Conv.