Improving the effectiveness of training the on-board object detection system for a compact unmanned aerial vehicle

Abstract

The model of object detector and the criterion of leaning effectiveness of the model were proposed. The model contains 7 first modules of the convolutional Squeezenet network, two convolutional multiscale layers and the information­extreme classifier. The multiplicative convolution of the particular criteria that takes into account the effectiveness of detection of objects in the image and accuracy of the classification analysis was considered as the criterion of learning effectiveness of the model. In this case, additional use of the orthogonal matching pursuit algorithm in calculating high­level features makes it possible to increase the accuracy of the model by 4 %. The training algorithm of object detector under conditions of a small size of labeled training datasets and limited computing resources available on board of a compact unmanned aerial vehicle was developed. The essence of the algorithm is to adapt the high­level layers of the model to the domain application area, based on the algorithms of growing sparse coding neural gas and simulated annealing. Unsupervised learning of high­level layers makes it possible to use effectively the unlabeled datasets from the domain area and determine the required number of neurons. It is shown that in the absence of fine tuning of convolutional layers, 69 % detection of objects in the images of the test dataset Inria Aerial Image was ensured. In this case, after fine tuning based on the simulated annealing algorithm, 95 % detection of the objects in test images is ensured. It was shown that the use of unsupervised pretraining makes it possible to increase the generalizing ability of decision rules and to accelerate the iteration process of finding the global maximum during supervised learning on the dataset of limited size. In this case, the overfitting effect is eliminated by optimal selection of the value of hyperparameter, characterizing the measure of coverage of the input data of by network neurons.