مجله رایانش نرم و فناوری اطلاعات
سال سیزدهم شماره 3 (پاییز 1403)

Handwriting recognition has been a highly challenging area in computer vision for the past twenty years. The main source of data for such systems is handwriting images, leading to the proposal of various architectures based on the availability of extensive labeled data. This study introduces a hybrid architecture called GCOCR, which utilizes deep learning methods and gated convolutional recurrent neural networks to intelligently recognize characters. The architecture incorporates pre-processing techniques such as padding, deslanting, and data augmentation. Additionally, it implements gated convolutional layers, ReLU activation functions, pooling, and a bidirectional neural network. The performance results of this architecture on the IAM dataset indicate that the character error rate (CER) of GCOCR is 8.6% lower and the word error rate (WER) of GCOCR is 2.4% lower than the strongest comparable research studies.

 One of the main challenges in machine learning is the need for extensive training data, which increases the cost and time required for data preparation. In some cases, there may be limited data available, or privacy concerns may arise. Generalized zero-shot learning (GZSL) offers a suitable option in such situations, enabling the prediction of both seen and unseen classes. Although various methods have been proposed in this field, accurately recognizing seen and unseen classes still presents multiple challenges. This paper introduces a hybrid generalized zero-shot learning (H-GZSL) approach, which improved classification accuracy in GZSL. The proposed method simultaneously utilizes two classification techniques, enhancing classification accuracy for both seen and unseen classes. To address different scenarios and resolve conflicts in classification, a correction mechanism is employed. This mechanism compares the predicted classes from the two methods and assigns the instance to the most appropriate one. The proposed method was evaluated on the AWA2, SUN, aPY, CUB, and FLOW datasets. The evaluation separately measured the performance on seen and unseen classes, and the harmonic mean of classification accuracy across these classes was calculated. Compared to previous methods, the proposed approach achieved an improvement in the harmonic mean accuracy on the CUB and SUN datasets.

The covid-19 is a new virus that causes upper respiratory system and lung infections. Radiographic imaging is used to monitoring various lung diseases and has recently been used to monitoring covid-19 disease, which to detect early and control the outbreak of the disease. Most research in this field has been devoted to articles based on deep learning methods. Due to the high training and testing time of deep learning-based models, in this paper a mixture of ELM based experts with trainable gating network (MEETG) is used. In MEETG, the advantages of ELM are used for designing the structure of ME. The ELM learning process is faster than SVM and MLP and has a better generalizability than them. In continuation, ILQP has been used as texture descriptor for feature extraction and MEETG has been used as a classifier to diagnose six types of lung disease and healthy lung. To evaluate the proposed model, from the RYDLS-20 dataset with unbalanced samples containing chest X-ray images belonging to seven classes including six different viral pneumonia disease lung diseases (Covid-19, SARS, MERS and varicella), bacteria (Streptococcus), fungus (Pneumocystis) and healthy lungs. Our main aim is to achieve the best possible identification for Covid-19 among other types of pneumonia and healthy lungs. Evaluation measures consisting of accuracy, precision, recall, and F-Score have been used to evaluate the proposed model. The experimental results revealed that F-Score for detection of Covid-19 is equal to 0.897 and the average of the total F-Score and the classification accuracy is 0.80 and 97.07%, respectively. The average of total F-Score of the proposed method improves 24%, 43%, 37% and 20% compared with KNN, DT, MLP and ELM, and compared with ensemble learning based method such as ME and bagging is about 19%.