Hi Xavier,
I see. I think the best solution out there for you is to use a pretrained-model out there, trained on ImageNet. It stores the weights of the convolution filters of the models trained on large datasets. So all you have to do is load the weights given in these numpy arrays into TensorFlow variables. You can find such a model here (https://www.cs.toronto.edu/~frossard/post/vgg16/), but not limited to this. There are several other popular models such as, AlexNet, ResNet, etc. Hope this helps