Inception has been introduced in Rethinking the Inception Architecture for Computer Vision, Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, Zbigniew Wojna, 2015, https://arxiv.org/abs/1512.00567 and the key idea is to have multiple sizes of convolutions acting as feature extraction and computing 1×1, 3×3, and 5×5 convolutions within the same module. The outputs of these filters are then stacked along the channel dimension and sent into the next layer in the network. This is described in the following image:

Inception-v3 is described in Rethinking the Inception Architecture for Computer Vision, while Inception-v4 is described in Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning, Christian Szegedy, Sergey Ioffe, Vincent Vanhoucke, Alex Alemi, 2016, https://arxiv.org/abs/1602.07261.