Line detection is an important computer vision task traditionally solved by Hough Transform. With the advance of deep learning, however, trainable approaches to line detection became popular. In this paper we propose a lightweight CNN for line detection with an embedded parameter-free Hough layer, which allows the network neurons to have global strip-like receptive fields. We argue that traditional convolutional networks have two inherent problems when applied to the task of line detection and show how insertion of a Hough layer into the network solves them. Additionally, we point out some major inconsistencies in the current datasets used for line detection.
Most modern convolutional neural networks (CNNs) are compute-intensive, making them infeasible to use in mobile or embedded devices. One of the approaches to this problem is to modify a usual deep CNN with shallow early-exit branches, appended to some convolutional layers . This modification, named BranchyNet, allows to process simple input samples without performing full volume of calculations, providing a speed-up on average. In this work we consider the problem of training a BranchyNet. We exploit a cascade loss function , which explicitly regularizes CNN’s average computation time, and modify it to use the entropy of branches’ prediction as confidence measure. We show, that on CIFAR10 dataset the proposed loss function provides a actual speed-up increase from 43% to 47% without quality degradation, comparing with the original loss function.
With the development of Artificial Neural Networks (ANNs), they are becoming key components in many computer vision systems. However, to train ANNs or other machine learning programs it is necessary to create large and representative datasets, which can be a costly, hard and sometimes even impossible task. Another important problem with such programs is the data drift: in real-world applications input data can change with time, and the quality of a machine learning system trained on the fixed dataset may deteriorate. To combat these problems, we propose a model of ANN-based machine learning classification system that can be trained during its exploitation. The system both classifies input examples and performs training on the data gathered during its operation. We assume that besides ANN there is an external module in the system that can estimate confidence of the answers given by ANN. In this paper we consider two examples of such external module: a separate, uncorrelated classifier and a module that estimates ANN output by searching recognized words in a dictionary. We conduct numerical experiments to study the properties of the proposed system and compare it to ANNs trained offline.
In this paper we study the problem of combining UAV obtained optical data and a coastal vector map in absence of satellite navigation data. The method is based on presenting the territory as a set of segments produced by color-texture image segmentation. We then find such geometric transform which gives the best match between these segments and land and water areas of the georeferenced vector map. We calculate transform consisting of an arbitrary shift relatively to the vector map and bound rotation and scaling. These parameters are estimated using the RANSAC algorithm which matches the segments contours and the contours of land and water areas of the vector map. To implement this matching we suggest computing shape descriptors robust to rotation and scaling. We performed numerical experiments demonstrating the practical applicability of the proposed method.