Dynamic Patch Convolution (DPConv)
Shuchao Deng · Journal of Physics Conference Series · 2021
Abstract Lightweight Convolutional Neural Networks (CNNs) due to the small amount of calculation and performance degradation, budget constraints depth (number of convolution layers), and width of CNN (number of channels), resulting in limited representation capabilities. To solve this problem, this paper proposes a new dynamic block convolution design to increase the depth or width of the model complexity without increasing the network. Instead of using a single convolution kernel in each layer, dynamic block convolution dynamically aggregates multiple parallel convolution kernels based on input-related attention. Assembling multiple kernels is not only computationally efficient, but also due to the small size of the kernel, but also has more representation capabilities. These kernels are aggregated in a non-linear manner through attention. The two methods of spatial dimension block and multi-head are used in the traditional CNN attention design. The image is divided into small images and then multiple attention mechanism weights are used for each small image. Compared with the traditional CNN attention, the experimental effect is better.