While a large convolution kernel can solve the problem of the receptive field, it increases model parameters, meaning that the real-time requirements cannot be met. For resolution issues, low-stage features have more detailed information, while the high-stage features have higher semantic information...