CNN-RDM: a new image processing model for improving the structure of deep learning based on representational dissimilarity matrix

• Published in: The Journal of supercomputing 2023-03, Vol.79 (4), p.4266-4290
• Main Authors: Heydaran Daroogheh Amnyieh, Zahra, Rastegar Fatemi, Seyed Mohammad Jalal, Rastgarpour, Maryam, Aghaee Ghazvini, Golnaz
• Format: Article
• Language: English
• Subjects: Artificial neural networks; Compilers; Computer Science; Image classification; Image processing; Interpreters; Machine learning; Neural networks; Processor Architectures; Programming Languages; Training
• ISSN: 0920-8542; 1573-0484
• DOI: 10.1007/s11227-022-04661-7

Convolutional neural networks (CNNs) are widely used to categorize images. Successful training of a CNN requires rapid convergence of its weights, which increases the efficiency of training. In this paper, a CNN-RDM model, based on CNN and representational dissimilarity matrix (RDM), is proposed. In the proposed model, a loss function is defined in the RDM whose output is a minimized dissimilarity matrix. Therefore, by combining this minimized matrix of RDM loss function with CNN cross-entropy, the final output is minimized. The inputs of CNN-RDM model are pixels with low-level features, and the output is a set of features. The CNN-RDM model can actually change the structure of the neural network inspired by the visual pathway. The performance of our model is evaluated using 50 batches, including 10 image classes, each containing 5 samples on the data sets Cifar10, Cifar100, and Coco. Evaluation results demonstrate that our CNN-RDM model has achieved an accuracy of 60% in image classification, which is superior to 51% accuracy of the obtained model.