الفهرس | Only 14 pages are availabe for public view |
Abstract Object recognition is a subject that is becoming more important in both industry and academia. The ability to recognize objects is a key problem for many intelligent applications. Developing an effective feature extraction method is one of the most difficult issues in object recognition. For this process, a variety of algorithms were developed, including self-organizing maps (SOMs), support vector machines (SVM), principal component analysis (PCA), and modern deep learning (DL) techniques, particularly convolutional neural networks (CNNs) and U-net. In this thesis, DL was employed to recognize faces, Arabic and English digits, as well as a variety of other objects. The use of DL for health sciences applications like COVID-19 infection prediction is also discussed. The main objective of this thesis is to propose a new optimized technique to improve the performance of the CNNs in recognizing objects. This objective is a achieved by four proposed techniques. In the first technique, Cyclic SOMs is proposed to improve the SOMs’ work as feature extractors. In the second technique, CNNs were enhanced using the self-organizing maps (SOMs) topology space in the convolution layer and the KNN classifier instead of the conventional fully connected layer. The third technique employs the KNN classifier in the fully connected layer and Cyclic-SOMs in the convolution layer of the CNNs. The fourth technique involves using DL in the healthcare sector to predict COVID-19 infections using U-net. The efficiency of the first three techniques has been evaluated on four wide benchmark datasets: AHDBase for Arabic digits, MNIST for English digits, CMUPIE for faces, and CIFAR-10 for objects. The experiment results of the proposed techniques on the four mentioned datasets provide the following findings in comparison to other techniques in terms of the recognition rate accuracy: the first technique produced accuracy 80.9%, 90.35%, 96.42%, and 85.77%. The second technique produced 96.57%, 95.4%, 97%, and 89.23%; while the third technique produced accuracy 97.7%, 98.2%, 98.51%, and 93.8%. Regarding the fourth technique, several evaluation metrics using CT scan dataset for the lung were used to measure the performance of our proposed algorithm in comparison with other state-of-the-art methods in terms of accuracy, sensitivity, precision, and dice coefficient. The experimental results of the proposed technique reached 99.71%, 0.83, 0.87, and 0.85, respectively, in comparison with existing techniques. |