Keywords: Siamese Neural Network, convolutional neural network, multi-directional image projections, radon transform, neural architecture generation..

Data preprocessing is a crucial task of data scientists, as well as lowering dimensionality of input vectors as possible. While deep learning got popular with big data used in neural networks with a large number of neurons, researchers are also working on methods where accurate predictions can be produced with less memory cost, efficiently. Nowadays machine learning algorithms are applied on devices with less computational power, smartphones, and IoT-driven smart cameras: in this case, the models should be small, as they have to fit into the memory of the device. In special cases of problems, e.g. images, Convolutional Neural Networks (CNN) are successfully applied, with relatively small trainable parameter numbers. However, as the input image size increases, the number of hidden layers must increase as well. Deep learning is effective in multi-class image classification and object detection with the use of end-to-end learning: in these cases, the input is not preprocessed, the input of the network is the data in its raw form of pixels. If there are fewer samples, well-known methods like feature elimination, feature selection or principal component analysis could be applied, where insignificant parts are simply ignored. In the case of images, of course, to decrease the dimensionality resizing or miniaturization can be applied as well.

A special type of classification is comparation, where the model must decide from two inputs that they belong to the same instance, or not. These type of problems are in most cases identification or re-identification of objects, for example, images of vehicles. In the case of image comparison, a special structure of two CNNs, with shared parameters is used, called a Siamese Neural Network (SNN). The paper introduces a method for image preprocessing based on a signature calculated from multi-directional projections. This method produces a smaller tensor than the original image itself, resulting in shallow CNN can be applied instead of the end-to-end solution. To analyze the results, multiple convolutional neural architectures are generated based on the input dimensionality and a memory limit using a novel method, and the architectures are then trained using a cluster of computers with graphical processors. For effective training of the models, an auto load-balancing system is designed and presented, along with the measured runtimes and comparative analysis of the results.

purchase the full-text of this paper (price £22)

go to the previous paper
go to the next paper
return to the table of contents
return to the book description
purchase this book (price £25 +P&P)