Figure 2
The figure is a left-to-right C N N architecture diagram illustrating feature extraction using depthwise separable convolution followed by pointwise convolution. On the far left, four small skin-related example images are grouped under the label “Preprosed images” and point via an arrow to a stack of three colored feature-map layers. From this stack, three parallel paths branch upward, middle, and downward into depthwise separable convolution operations. The upper path shows a single green feature map passing through a block labeled “Depthwise Separable Conv”. The middle path shows a red feature map passing through a block labeled “D subscript k by D subscript k Conv”. The lower path, followed by ellipsis, shows a blue feature map passing through another depthwise convolution block. The outputs of these depthwise operations are shown as gray feature-map blocks, which then merge into a combined stack of three gray layers. An arrow leads to the next stage labeled “Pointwise Conv” with “1 by 1 Conv”, producing another stack of gray feature maps. The final arrow points to a vertical feature vector illustration on the right, enclosed in a dotted rectangle and labeled “1,280 extracted features”, indicating the size of the extracted feature representation.A proposed lightweight depthwise separable CNN (LDSCNN) architecture for efficient feature extraction from dermatological images