Search

Predicting the sequence pattern of the handwritten text images is a challenging problem due to various writing styles, insufficient training data, and also background noise appearing in the text images. The architecture of the combination between convolutional neural network (CNN) and recurrent neural network (RNN), called CRNN architecture, is the most successful sequence learning method for handwritten text recognition systems. For handwritten text recognition in historical Thai document images, we first trained nine different CRNN architectures with both training from scratch and transfer learning techniques to find out the most powerful technique. We discovered that the transfer learning technique does not significantly outperform scratch learning. Second, we examined training the CRNN model by applying the basic transformation data augmentation techniques: shifting, rotation, and shearing. Indeed, the data augmentation techniques provided more accurate performance than without applying data augmentation techniques. However, it did not show significant results. The original training strategy aimed to find the global minima value and not always solve the overfitting problems. Third, we proposed a cyclical data augmentation strategy, called CycleAugment, to discover many local minima values and prevent overfitting. In each cycle, it rapidly decreased the training loss to reach the local minima. The CycleAugment strategy allowed the CRNN model to learn the input images with and without applying data augmentation techniques to learn from many input patterns. Hence, the CycleAugment strategy consistently achieved the best performance when compared with other strategies. Finally, we prevented image distortion by applying a simple technique to the short word images and achieved better performance on the historical Thai document image dataset.

There is a global increase in health awareness. The awareness of changing eating habits and choosing foods wisely are key factors that make for a healthy life. In order to design a food image recognition system, many food images were captured from a mobile device but sometimes include non-food objects such as people, cutlery, and even food decoration styles, called noise food images. These issues decreased the performance of the system. Convolutional neural network (CNN) architectures are proposed to address this issue and obtain good performance. In this study, we proposed to use the ResNet50-LSTM network to improve the efficiency of the food image recognition system. The state-of-the-art ResNet architecture was invented to extract the robust features from food images and was employed as the input data for the Conv1D combined with a long short-term memory (LSTM) network called Conv1D-LSTM. Then, the output of the LSTM was assigned to the global average pooling layer before passing to the softmax function to create a probability distribution. While training the CNN model, mixed data augmentation techniques were applied and increased by 0.6%. The results showed that the ResNet50+Conv1D-LSTM network outperformed the previous works on the Food-101 dataset. The best performance of the ResNet50+Conv1D-LSTM network achieved an accuracy of 90.87%.

Use of ensemble convolutional neural networks (CNNs) has become a more robust strategy to improve image classification performance. However, the success of the ensemble method depends on appropriately selecting the optimal weighted parameters. This paper aims to automatically optimize the weighted parameters using the differential evolution (DE) algorithm. The DE algorithm is applied to the weighted parameters and then assigning the optimal weighted to the ensemble method and stacked ensemble method. For the ensemble method, the weighted average ensemble method is applied. For the stacked ensemble method, we use the support vector machine for the second-level classifier. In the experiments, firstly, we experimented with discovering the baseline CNN models and found the best models on the pornographic image dataset were NASNetLarge with an accuracy of 93.63%. Additionally, three CNN models, including EfficientNetB1, InceptionResNetV2, and MobileNetV2, also obtained an accuracy above 92%. Secondly, we generated two ensemble CNN frameworks; the ensemble learning method, called Ensemble-CNN and the stacked ensemble learning method, called StackedEnsemble-CNN. In the framework, we optimized the weighted parameter using the DE algorithm with six mutation strategies containing rand/1, rand/2, best/1, best/2, current to best/1, and random to best/1. Therefore, the optimal weighted was given to classify using ensemble and stacked ensemble methods. The result showed that the Ensemble-3CNN and StackedEnsemble-3CNN, when optimized using the best/2 mutation strategy, surpassed other mutation strategies with an accuracy of 96.83%. The results indicated that we could create the learning method framework with only 3 CNN models, including NASNetLarge, EfficientNetB1, and InceptionResNetV2.