Dynamic Stride Convolution and Its Inter-Layer Interpretable Method
The image processing schemes based on convolutional neural network sets the convolution step to a fixed value independent of the input image.The equal convolution allocating for both important and unimportant areas of input images leads to unreasonable resource allocation and network redundancy.To address this problem,we propose dynamic stride convolution(DSC),which modifies convolution strides of the convolution kernel by learning a set of offsets related to the input data,and allocates more computations adaptively to the regions of interests.Furthermore,an inter-layer interpreta-ble method is proposed to visualize the convolution distribution using the learned offset,which can generate intuitive inter-pretable diagram with very low computational consumption and help researchers analyze the attention distribution between the convolutional layers.In order to further optimize the convolutional resource allocation,a new loss function is designed to effectively improve the performance of DSC and achieve the editing of resource locations,and the inter-layer interpreta-ble analysis method is combined to visualize resource editing.DSC is embedded into different tasks such as object detection and image segmentation,and experimental results show that the mAP of different networks on the COCO datasets have in-creased by more than two percents generally,which shows the effectiveness of DSC method.
万方数据
维普
