Image registration is a typical problem and technical difficulty in the research field of image processing, which can be used in infrared image fusion, image Mosaic, image segmentation, super-resolution reconstruction and other directions. On the basis of infrared image registration, considering the difference of infrared image, the effective accuracy of registration is set. The algorithm in this paper was initially used in super-resolution reconstruction, but it was redesigned in the registration process due to the need to reduce the computing cost in hardware implementation. The algorithm is improved on the basis of the original flownetS, and the cost of the calculate is reduced by 89% when the number of convolutions layers and the receptive field of the improved network are the same. There is no need for pooling layer, because pooling provides greater receptive field while reducing resolution, which will lead to spatial information loss and data loss. Dilate convolution convolution can avoid down-sampling and provide a larger sense field under the premise of the same amount of computation. Set different Dilation rates, receptive fields will be different, and there will be multi-scale information. This method based on deep learning uses the processing method of dilated convolution to greatly reduce the cost of network computing and provide a convenient channel for the hardware implementation of the algorithm. Moreover, the algorithm also achieves excellent results on the development board RK3399Pro. Combined with infrared image, this paper also demonstrates the difference between infrared image in the registration of effective accuracy and visible light, and also analyzes the target at different speed, the minimum network structure required. Finally, an appropriate ratio was selected through experimental attempts to ensure stable accuracy in the conventional moving target.
|