Deep neural networks (DNNs) have made outstanding achievements in a wide variety of domains. For deep learning tasks, large enough datasets are required for training efficient DNN models. However, big datasets are not always available, and they are costly to build. Therefore, balanced solutions for DNN model efficiency and training data size have caught the attention of researchers recently. Transfer learning techniques are the most common for this. In transfer learning, a DNN model is pre-trained on a large enough dataset and then applied to a new task with modest data. This fine-tuning process yields another challenge, named catastrophic forgetting. However, it can be reduced using a reasonable strategy for data argumentation in incremental learning. In this paper, we propose an efficient solution for the random selection of samples from the old task to be incrementally stored for learning a sequence of new tasks. In addition, a loss combination strategy is also proposed for optimizing incremental learning. The proposed solutions are evaluated on standard datasets with two scenarios of incremental fine-tuning: (1) New Class (NC) dataset; (2) New Class and new Instance (NCI) dataset. The experimental results show that our proposed solution achieves outstanding results compared with other SOTA rehearsal methods, as well as traditional fine-tuning solutions, ranging from 1% to 16% in recognition accuracy.

In this study, we extensively analyze and evaluate the performance of recent deep neural networks (DNNs) for hand gesture recognition and static gestures in particular. To this end, we captured an unconstrained hand dataset with complex appearances, shapes, scales, backgrounds, and viewpoints. We then deployed some new trending convolution neuron networks (CNNs) for gesture classification. We arrived at three major conclusions: i) DenseNet121 architecture is the best recognition rate through almost evaluated red, green, blue (RGB) and augmentation datasets. Its performance is outstanding in most original works; ii) blender-based augmentation help to significantly increase 9% of accuracy, compared to the use of a RGB cues; iii) most CNNs can achieve impressive results at 97% accuracy when the training and testing datasets come from the same lab-based or constrained environment. Their performance is drastically reduced when dealing with gestures collected in unconstrained environments. In particular, we validated the best CNN on a new unconstrained dataset. We observed a significant reduction with an accuracy of only 74.55%. This performance can be improved up to 80.59% by strategies such as blender-based and/or GAN-based data augmentations to obtain an acceptable result of 83.17%. These findings contribute crucial factors and make fruitful recommendations for the development of a robust hand-based interface in practice

Convolution Neural Networks (CNNs) have become an attractive method for the detection of anomalous behaviors. However, designing an efficient CNN model in terms of classification accuracy remains a challenging problem. Furthermore, the existing datasets for abnormal behavior detection are limited, with each focusing on a certain context. Therefore, a CNN model trained on a certain dataset will be adaptive for a particular context and not suitable for other contexts. This study proposes a CNN framework with an efficient attention mechanism to capture key information from multiple inputs, namely RGB, optical flow, and heatmap. Experiments were carried out on several benchmark datasets and a self-collected dataset, and the evaluation involved both single- and cross-dataset strategies. The results show the superior performance of the proposed frameworks compared to other SOTA methods in detection accuracy.