To create 64 Gabor features at each scale and orientation in the spatial and frequency domain, you can follow these steps:

1. Define the range of scales and orientations for which you want to create the Gabor filters. For example, you could define 8 scales and 8 orientations.

2. Determine the size of the filter kernels that you want to use. The size of the kernel will depend on the scale of the filter, with larger filters being used for coarser scales. For example, you might use a 7x7 kernel for the coarsest scale and a 19x19 kernel for the finest scale.

3. Choose the frequency of the filter. The frequency of the filter will determine how many oscillations the filter has per unit distance. This will affect the size and shape of the filter kernel.

4. Create a set of Gabor filters for each scale and orientation. To do this, you can use the following formula:

   G(x,y) = exp(-((x^2 + y^2)/(2sigma^2))) * cos(2pif(xcos(theta) + ysin(theta)) + phi)

   where G(x,y) is the filter kernel, sigma is the standard deviation of the Gaussian envelope, f is the frequency of the sinusoidal carrier, theta is the orientation of the filter, and phi is the phase offset of the filter.

5. Apply each filter to the input image in both the spatial and frequency domains to obtain the corresponding Gabor feature. In the spatial domain, this involves convolving the filter kernel with the input image. In the frequency domain, this involves multiplying the Fourier transform of the filter kernel with the Fourier transform of the input image, and then taking the inverse Fourier transform of the result.

6. Repeat this process for each scale and orientation to obtain a set of 64 Gabor features at each scale and orientation in both the spatial and frequency domains.

Note that there are libraries available in various programming languages such as Python, Matlab, and C++ that can be used to create Gabor filters and extract Gabor features from images. These libraries can simplify the implementation process and provide additional functionality, such as the ability to perform feature normalization and pooling.