Diabetes mellitus (DM) is one of the major diseases that cause death worldwide and lead to complications of diabetic foot ulcers (DFU). Improper and late handling of a diabetic foot patient can result in an amputation of the patient’s foot. Early detection of DFU symptoms can be observed using thermal imaging with a computer-assisted classifier. Previous study of DFU detection using thermal image only achieved 97% of accuracy, and it has to be improved. This article proposes a novel framework for DFU classification based on thermal imaging using deep neural networks and decision fusion. Here, decision fusion combines the classification result from a parallel classifier. We used the convolutional neural network (CNN) model of ShuffleNet and MobileNetV2 as the baseline classifier. In developing the classifier model, firstly, the MobileNetV2 and ShuffleNet were trained using plantar thermogram datasets. Then, the classification results of those two models were fused using a novel decision fusion method to increase the accuracy rate. The proposed framework achieved 100% accuracy in classifying the DFU thermal images in binary classes of positive and negative cases. The accuracy of the proposed Decision Fusion (DF) was increased by about 3.4% from baseline ShuffleNet and MobileNetV2. Overall, the proposed framework outperformed in classifying the images compared with the state-of-the-art deep learning and the traditional machine-learning-based classifier.

This paper discusses the initial experiments on fine-tuning the convolutional neural network (CNN) models of ResNet101, DenseNet, MobileNetV2, and ShuffleNetV2 for breast cancer detection. These models have been tested using the ImageNet database and shown their high performance on the image classification. Here, the dataset used to train the model are thermal breast images downloaded from the Database for Mastology Research (DMR). There are only two classification classes here, cancer and healthy. During the training, we used three epochs for the training iteration, 10, 20, and 30. For this preliminary study, we set the learning rate to 0.001, momentum to 0.9, learning rate factor for weight and bias each to 10, and minibatch size to 10. The training results showed that ResNet101 and DenseNet with deep networks have 100% accuracy only in 10 epochs. Whereas, MobileNetV2 and ShuffleNetV2 respectively need 20 epochs and 30 epochs of training to achieve 100% accuracy. Finally, we evaluated the performance of each pretrained model using testing results. The DenseNet was able to classify all the testing dataset correctly. ResNet101 and MobileNetV2 have the same performance, correctly classified static dataset while slightly missed in classifying dynamic dataset with 0.996 of accuracy. ShuffleNetV2 has a little lower performance with only 0.98 of accuracy. In term of training time, ShuffleNetV2 spent a short duration, but MobileNetV2 with competitive elapse time was able to show equal performance to ResNet101.

Document image binarization is a challenging task because of combined degradation in a document. In this study, a new binarization method is proposed for binarizing an ancient document with combined degradation. The proposed method comprises the following four stages: histogram analysis, contrast enhancement, local adaptive thresholding, and artifact removal. In histogram analysis, a new approach is applied to establish a uniform background. Next, the image contrast is enhanced using a new contrast enhancement, and then the document is binarized using a novel local adaptive thresholding. Artifacts from the binarization process are removed in the artifact removal stage. Finally, an experiment is conducted using one private and four public datasets and by simulating the proposed method with and without contrast enhancement. The results showed that the proposed method is faster and more effective compared to other state-of-the-art procedures for binarizing ancient documents.

Most local adaptive image binarization techniques have been inspired by Niblack's method and thus use local thresholding. Niblack's and the NICK methods that is based on it introduce a parameter k to determine the object boundaries in a given window, but one of the deficiencies of local thresholding is that the k value is constant. In this paper, we propose a new approach to calculating k values for the NICK binarization method: we set the k value based on the standard deviation of the image. For our experiments, we used the DIBCO 2013 dataset and a privately held ancient Jawi manuscript dataset. The results were evaluated using the F-measure, pseudo F-measure, peak signal-to-noise ratio, misclassication penalty metrics (MPM), distance reciprocal distortion (DRD), and overall rank score. The proposed method achieved result of 91.09% for the DIBCO dataset and 87.07% for the Jawi dataset, which were higher than those with earlier methods that used fixed k values ranging from - 0.2 to - 0.1. These results indicate that the k values produced by the proposed method can adapt to the state of the manuscript and that using them for NICK thresholding can increase binarization performance.

Binarization is one of the most critical steps in document image analysis, especially in Optical Character Recognition (OCR) system. Binarization process results in two image representations: black pixel represents the text or object, and white pixel represents the background. Thresholding is the simplest binarization approach. In this paper, we proposed a new adaptive thresholding technique for binarizing ancient document. The proposed method is the combination of Niblack and Sauvola techniques. It solved Niblack and Sauvola drawbacks by determining the mean of the thresholding values of both methods. The binarization results were evaluated using FM, FMps, PSNR, NRM, DRD, and MPM. The results showed that the proposed method had remarkable performance in binarizing ancient document, especially ancient Jawi document.