This network employs skip connections like ResNet architecture in contrast to version 1. It helps the model in faster training and to build deeper networks. There are also linear bottlenecks present in the middle of layers. It helps in two ways as we reduce the number of input channels:

1. With depthwise convolution, the model size is reduced
2. At inference time, memory usage is low. V2 denotes version 2 of mobilenet and 1.0X uses the original parameter settings.

More details can be found here.
Left: Standard convolutional layer with batchnorm and ReLU. Right: Depthwise Separable convolutions with Depthwise and Pointwise layers followed by batchnorm and ReLU.