OrthoMaps: an efficient convolutional neural network with orthogonal feature maps for tiny image classification. Issue 12 (18th September 2019)

• Record Type: Journal Article
• Authors: Moradi, Reza; Berangi, Reza; Minaei, Behrooz
• Is Part Of: IET image processing. Volume 13:Issue 12(2019)
• Page Start: 2067
• Page End: 2076
• Publication Date: 2019-09-18
• Subjects: learning (artificial intelligence) -- visual databases -- image classification -- feature extraction -- convolutional neural nets; OrthoMaps -- orthogonal feature maps -- tiny image classification -- image processing domain -- deep learning -- visual data -- learnable variables -- training process -- memory resources -- residual modules -- model architecture -- memory costs -- additive penalty -- objective function -- elicit uncorrelated representations -- input images -- deeper networks -- informative representation -- periodic learning rate curves -- total cost -- input data -- convolutional neural network -- fractional pooling; Image processing -- Periodicals; 621.36705
• DOI: 10.1049/iet-ipr.2018.6620
• Languages: English
• ISSNs: 1751-9659

Abstract

Abstract : In image processing domain of deep learning, the big size and complexity of the visual data require a large number of learnable variables. Subsequently, the training process consumes enormous computation and memory resources. Based on residual modules, the authors developed a new model architecture that has a minimal number of parameters and layers that enabled us to classify tiny images using much less computation and memory costs. Also, the summation of correlations between pairs of feature maps as an additive penalty in the objective function was used. This technique encourages the kernels to be learned in a way that elicit uncorrelated representations from the input images. Also, employing Fractional pooling helped to have deeper networks that consequently resulted in more informative representation. Moreover, employing periodic learning rate curves, multiple machines are trained with a less total cost. In the training phase, a random augmentation to the input data that prevent the model from being overfitted was applied. Applying MNIST and CIFAR‐10 datasets to the proposed model resulted in the classification accuracy of 99.72 and 93.98, respectively.