GNGTS 2023 - Atti del 41° Convegno Nazionale

Session 3.1 GNGTS 2023 network with different filter dimensions, to extract information on the input at different resolutions (by downscaling and upscaling the size of the layer input, respectively). Subsequently, the output is obtained after concatenating the feature maps extracted from each downscaling layer with the feature maps extracted from the corresponding upscaling layer (Fig.1). For the purpose of comparing results in terms of prediction time and quality, several learning strategies were analysed: Curriculum Learning and single-step learning. The former refers to a strategy introduced by Bengio et al., 2009, in which the model undergoes a pretraining phase on a large number of simple training data and, at a later stage, the model undergoes fine-tuning based on fitting the pre-trained model on a smaller number of but more complex data, following the idea (verified for a wide variety of applications) that learning based on a gradual increase in the complexity of the examples (Fig.2) brings the model back to a more efficient optimisation. In this work, complexity of the instances coincides with the features of synthetic LWD images, in terms of the number of discontinuities in the image, noise thresholds and possible intersections. All listed complexity factors were chosen based on human experience and in the synthetic generation stages are controlled by randomness components. In the single-step case, on the other hand, the complexity of the training examples is given in random order. For both strategies, 10 % of 1’000’000 training examples were used as a set of values. Figure 1: U-Net architecture scheme: it’ s a symmetric architecture composed by a down-sampling (yellow arrows) and an up-sampling path (grey arrows), that can capture high resolution features and context, respectively. The concatenation (black arrows) of the results of up-sampling and down-sampling layers provides more precise segmentation outputs in terms of localization, as it was demonstrated by Ronnenberger et al., 2015.