One way of increasing the efficiency of the polymer electrolyte membrane is to study two-phase flow effects in order to improve the evacuation of the product of the reaction, namely the water in fuel cells and the dioxygen in the electrolyser. The analysis of the water evacuation in PEMFC has been studied extensively, but the formation and the evacuation of dioxygen in PEMWE are not yet systematic. We propose to detect water bubbles through 3 actions in order to observe and identify dioxygen bubbles. Firstly, the image segmentation via morphological processing to create a sketchily data set. This approach has some limits, such as size or shape detection. Therefore, for a second stage a training of a shape regression convolutional neural network is used to estimate shape and complex forms [4]. And finally, the last action a validation of the neural model in online and real-time is provided. We hope that this work will help further studies, allowing to link shape, frequency, residence time, and velocity of bubbles of the different parameters used with electrolyser.