Is there any publication or research article from Zimmerman et al. or any other scholar in which it is mentioned that using small kernal size i.e. 3x3 achieve better results?
You are probably referring to this article:
Arguably, their main message is not that small kernels are better, rather demonstrating that the depth of the network is much more important than the size of the kernels. Also, they accredit the idea of using small kernels to Cireşan, D. C. et al. (2011): Flexible, High Performance Convolutional Neural Networks for Image Classification, however, this work performs no thorough comparison of effects of using small kernel size.