The automated classification of physical violence in images faces a critical methodological obstacle: class imbalance, which undermines the discriminative capacity of Deep Learning (DL) models by leading them to overfit the majority class. Although numerous balancing strategies have been proposed in the literature, which range from traditional resampling to advanced generative models, a systematic and controlled empirical evaluation of their behavior in the specific domain of physical violence detection remains absent. In this study, we present a systematic empirical review that evaluates traditional balancing techniques, generative models, and data augmentation methods under controlled imbalance conditions. The methodology is applied to five datasets (UBI-FIGHTS, RLVSD, RWF-2000, RLVS, and AIRTLAB), across six imbalance scenarios defined by Balance Percentage levels of 1%, 5%, 10%, 25%, 50%, and 75%. To isolate the effect of the balancing strategies from architectural confounders, a single stable backbone (ResNet-18) is employed as a controlled classifier across all experimental conditions. The experimental results demonstrate the absence of a universally optimal balancing technique: in highly imbalanced scenarios, traditional oversampling methods such as ROS achieved the best average performance, while generative approaches such as DCGAN and DeepSMOTE became increasingly competitive as class balance is improved. These findings confirm that the effectiveness of balancing techniques depends on both the degree of asymmetry and the dataset’s intrinsic characteristics. Thus, it is clear that there is a need for context-aware strategy selection rather than one-size-fits-all solutions. Beyond the empirical findings, this work provides a structured synthesis of the theoretical foundations and state-of-the-art methods for imbalanced violence detection.

The detection of the faces of people in self-access systems, is a challenging task since the pandemic of the coronavirus (SARS-CoV2) has changed our way of living, especially because companies and some government entities such as hospitals need to grant permission to access their facilities especially if the person does not bring the mask, and although many facilities no longer allow its use, since the above mentioned entities if they justify its use especially if their facilities there are people with chronic diseases that need and justify the use of the mask. In this work, we mention an access system that consists of software embedded in a Rasp-berry PI device, which contains an application formed by the use of a neural network Multi-task cascade convolutional networks (MTCNN), with which the training of the images of faces that bring and do not bring mouthguards is done, Later, and for a better identification of the faces, an algorithm based on Local Binary Patterns Histogram (LBPH) is used, with which the characteristics of the face are obtained and later it is possible to classify if the person to be identified has or does not have a face mask. In the end, our proposed system has an average accuracy of 93% in the detection of faces with face masks.

Automated violence detection in images presents a technical and scientific challenge that demands specialized methods to enhance classification systems. This study introduces an approach for automatically identifying relevant samples to improve the performance of neural network models, specifically DenseNet121, with a focus on violence classification in images. The proposed methodology begins with an initial training phase using a balanced dataset (DS1, 6000 images). Based on the model’s output scores ($ou{t}_N$), three confidence levels are defined: Safe ($ou{t}_N\ge 0.9+\sigma$ or $ou{t}_N\le 0.1-\sigma$), Border ($0.5-\sigma \le ou{t}_N\le 0.5+\sigma$), and Average ($0.4-\sigma \le ou{t}_N\le 0.6+\sigma$). These levels correspond to scenarios with low, moderate, and high prediction error probabilities, respectively, where $\sigma$ is an adjustable threshold. The Border subset exhibits symmetry around the decision boundary ($ou{t}_N=0.5$), capturing maximally uncertain samples, while the Safe regions reflect functional asymmetries in high-confidence predictions. Subsequently, these thresholds are applied to a second dataset (DS2, 5600 images) to extract specialized subsets for retraining ($D{S}_{Safe}$, $D{S}_{Border}$, and $D{S}_{Average}$). Finally, the model is evaluated using an independent test set ($D{S}_{test}$, 4400 images), ensuring complete data isolation. The experimental results demonstrate that the confidence-based subsets offer competitive performance despite using significantly fewer samples. The Average subset achieved an F1-Score of 0.89 and a g-mean of 0.93 using only 20% of the data, making it a promising alternative for efficient training. These findings highlight that strategic sample selection based on confidence thresholds enables effective training with reduced data, offering a practical balance between performance and efficiency when symmetric uncertainty modeling is exploited.