Z-pooling**This research was supported by the program Cátedras Francesa do Estado de São Paulo, an initiative of the French consulate and the state of São Paulo (Brazil). The authors thank D. Fourer and I. Brahim for their contributions.

摘要

Deep neural networks (DNN) have revolutionized orthodox tasks of image analysis in, which they have accomplished outstanding results and continually do so. By employing modifications to the architectures and introducing various techniques (often greedy), considerable improvements have been achieved. We prove it in proposing a new pooling method based on Zeckendorf's number decomposition. The objective of Z pooling - as maximum pooling - is to sub-sample the input representation (image, hidden layer output matrix, etc.), by reducing its dimensionality and by making it possible to do hypotheses on the characteristics contained in the grouped sub-regions. But it is shown that Z pooling properties are better adapted to segmentation tasks than other pooling functions. The method was evaluated on a traditional image segmentation task, and on a dense labeling task carried out with a series of deep learning architectures.