Fine-tuning on neural nets?

For some reason I still don't understand the concept of fine tuning on neural networks. My main interest is in image classifcation, and when I read tutorials they mention that to make training easier, just use a architecture pretrained from imagenet, then just finetune using the existing architecture for your specific classification problem. Could someone explain what this means to me? Its probably really basic, I just can't get it. Thanks!