NIK Image recognition and Machine learning

Author and abstract
Hege Haavaldsen, Max Aasbø, Frank Lindseth and Håkon Hukkelås:
Autonomous Vehicle Control: End-to-end Learning in Simulated Environments
This paper examines end-to-end learning for autonomous vehicles in diverse, simulated environments containing other vehicles, traffic lights, and traffic signs; in weather conditions ranging from sunny to heavy rain. The paper proposes an architecture combing a traditional Convolutional Neural Network with a recurrent layer to facilitate the learning of both spatial and temporal relationships. Furthermore, the paper suggests a model that supports navigational input from the user to facilitate the use of a global route planner to achieve a more comprehensive system. The paper also explores some of the uncertainties regarding the implementation of end-to-end systems. Specifically, how a system’s overall performance is affected by the size of the training dataset, the allowed prediction frequency, and the number of hidden states in the system’s recurrent module. The proposed system is trained using expert driving data captured in various simulated settings and evaluated by its real-time driving performance in unseen simulated environments. The results of the paper indicate that end-to-end systems can operate autonomously in simulated environments, in a range of different weather conditions. Additionally, it was found that using ten hidden states for the system’s recurrent module was optimal. The results further show that the system was sensitive to small reductions in dataset size and that a prediction frequency of 15 Hz was required for the system to perform at its full potential.
Espen Myrum, Simen Andre Nørstebø, Sony George, Marius Pedersen and Jon Museth:
An automatic image-based system for detecting wild and stocked fish
Fish stocking is the method of raising fish in a hatchery and releasing them into a river or lake to sustain or increase an existing population or to create a population. This has been practised in many countries, including Norway. Before the fish are released, the adipose fin is commonly removed in order to identify that it is a stocked fish. Cameras have been mounted in several Norwegian rivers in order to monitor fish populations. Classification of fish from these cameras is today a manual task carried out by people. In this paper we propose an automatic classification method to separate wild fish from stocked fish using machine learning. Experiments on an image set of trouts (Salmo Trutta) show a very high accuracy of the proposed method.
Jørgen Bakløkken, Felix Schoeler, Hugo Nørholm, Sony George, Marius Pedersen and Børre Dervo:
Automated salamander recognition using deep neural networks and feature extraction
This paper presents a study conducted to recognize salamanders by using their unique body markings based on images. The detection and matching of unique patterns in a salamander’s body can be complex due variability in individual animals size, shape, orientation and also influence from the external enviornment. While traditional methods require time intensive manual image corrections of the salamanders to achieve accurate recognition, in this work we propose a fully automatic techinque for straigthening. We also propose a matching technique based on the corrected images. The convolutional neural network ResNet50 and dense scale-invariant feature transform (DSIFT) are used for belly pattern localization, and matching for salamander recognition.
Frode Tennebø and Marius Geitle:
Evaluating Population Based Training on Small Datasets
Recently, there has been an increased interest in using artificial neural networks in the severely resource-constrained devices found in Internet-of-Things networks, in order to perform actions learned from the raw sensor data gathered by these devices. Unfortunately, training neural networks to achieve optimal prediction accuracy requires tuning multiple hyper-parameters, a process which has traditionally taken many times the computation time of a single training run of the neural network. In this paper, we empirically evaluate the Population Based Training algorithm, a method which simultaneously both trains and tunes a neural network, on datasets of similar size to what we might encounter in an IoT scenario. We determine that the population based training algorithm achieves prediction accuracy comparable to a traditional grid or random search on small datasets, and achieves state-of-the-art results for the Biodeg dataset.