Segmentação de Imagens incluindo Contexto em Redes Neurais Convolucionais

Abstract

There is a significant demand for the automation of the location and recognition of objects and people, from the automation of agriculture to systems for automatic measurement of the water level in rivers, all performed by computer vision systems. These markings or labels are currently assigned at the pixel level, a technique called semantic segmentation. However, in a single image there can be several classes, and often these classes are very similar, making it a complex challenge to be worked on. Recently, methods based on Convolutional Neural Networks (CNN) have achieved impressive success in semantic segmentation tasks. This success is due, among other factors, to the inclusion of some context to assist the network, such as the information that one class is more frequent than the other and/or; the information that the dataset has images with a high level of pixel-labeling uncertainty present at the edges. However, these two points mentioned, both class imbalance and pixel-labeling uncertainty, can be further explored. We present an approach that calculates and assigns a pixel-wise weight, considering its class and the uncertainty during the labeling process. Pixel-wise weights are used during training to increase or decrease the importance of the pixels. Some papers are presented demonstrating the use of semantic segmentation techniques with context inclusion, with significant results in comparison with the most relevant methods. In addition, we also present a method for the reconstruction of the area of the object of interest, allowing the reconstruction of the edges of this object. The techniques presented here can be used in a wide variety of segmentation methods, improving their robustness.