In the visual cortex, the size of the local receptive field of neurons in the same area is different, so different sizes of spatial information can be obtained in the same processing stage. However, when designing convolution network, other properties of receptive field are not considered, such as adaptive adjustment of receptive field size. Visual cortex<br>The size of receptive field of neuron in the next convolution layer is stimulated and modulated. In ong, a network with multiple branches, there is a potential mechanism to adjust the Q size of receptive field of neuron in the next convolution layer according to the input content. This is because the next convolution layer fuses the characteristics of different branches through linear combination, However, this method provides a strong ability to adjust the network. Sknet is a non-linear method to fuse the features from different nuclei to achieve different size adjustment of receptive field. It includes three operations: SPLT operation generates multiple channels with different kernel sizes, which are related to different receptive field sizes of neurons; fuse operation combines the information from multiple channels to obtain a global and understandable representation for weight selection; and fuse operation combines the information from multiple channels to obtain a global and understandable representation for weight selection; The select operation fuses feature maps with different core sizes according to the selected weights.<br>
正在翻译中..