The proposed contribution describes a new convolutinal neural network structure that uses several kernels extracted from the F-transform methodology and shows the experimental results. The purpose of the ongoing study is to explore optimisation possibilities of the training process. We have performed tests using the newly proposed network model and the Caffe framework. The proposed version prototype is the network LeNet-5. We have passed multiple training on the data-set MNIST. The time measurements were taken in each training cycle. After the training stage, we have tested the accuracy and the robustness against rotation and translation of the proposed and the prototype networks. The median training time of the prototype network was 53.81 second while median training time of the proposed network was 37.27 second. Both networks had similar accuracy at the end of the training process (99.08% and 98.94%). The performance was similar on rotated example image. The performance on translated example image is going to be the subject of the further study. The results show that excluding some of the network parameters from training process and initialising them based on the F-transform lead to the significantly lower training time while retaining network accuracy.