Semi-automatic image generation for neural networks datasets

Machine Learning and Deep Learning algorithms require to gather a substantial amount of data to train on. Scarcity of relevant data can make it difficult to prepare networks efficiently, so the creation of useful data appears as an interesting path to explore. In our case, the sparse data that had to be generated were plastic parts with defects on them. In order to create synthetic irregularities, several individual defects were randomly picked, warped and placed on the pictures of flawless plastic pieces. A brief succession of experiments was carried out to create a script that would allow the generation of such data semi-automatically. The process was “semi-automatic” as flawless pieces had to be cut from their backgrounds, and defects from their surrounding textures, manually beforehand. Thanks to the mixture of data the scripts generate, datasets composed of artificial pictures can be used to train/test algorithms on data that would have been laborious to get otherwise. Experimentations proved that these synthetic data were viable to train CNNs and that these networks could process new data effectively..