Deep Learning (DL)for faces is a critical component in various applications, including surveillance, marketing, and security. However, when dealing with human subjects, it is often necessary to determine whether facial anonymization is required in order to preserve privacy. Traditional object detectors only provide bounding boxes and confidence scores, offering no information about the level of privacy risk associated with each detected face. To the best of our knowledge, this is the first DL approach that explicitly estimates the degree of privacy required for each face. Our proposed method extends conventional object detectors by introducing a novel Privacy Awareness metric, which quantifies the extent of privacy exposure for each detected face. In addition to the standard outputs —bounding box coordinates and confidence scores— our model provides this privacy-related information, enabling downstream systems to apply appropriate anonymization strategies. We present experimental results demonstrating that our approach not only maintains strong DL performance but also provides meaningful privacy assessments. Furthermore, we analyze the effectiveness of various combinations of Deep Learning (DL) models used in our framework, identifying configurations that offer the best performance for privacy-aware face detection.

Semi-supervised learning has gained significant attention recently, especially in the context of classification problems. This paper focuses on the object detection problem, which involves not only classification but also the localization of objects within an image—introducing new challenges when applying semi-supervised learning in this domain. Among semi-supervised solutions, the use of pseudo-labels is typically the go-to method. However, one of the main issues with pseudo-labels is the inaccuracy in object localization, which presents a specific challenge. While others have attempted to address the localization issue by modifying the main model, this paper proposes a novel approach that introduces a novel non-maximum suppression method. This method leverages additional information to improve the quality of pseudo-labels, thereby further optimizing the model’s performance. The experimental results demonstrate substantial improvements in performance when compared to well-established baseline models, highlighting the effectiveness of the proposed method.

Object detection (OD) typically demands large annotated datasets and substantial computational resources. To address these challenges, we propose a novel two-stage pipeline that integrates Visual Foundation Models (VFMs) for object proposal generation with few-shot learning models enhanced by Mahalanobis distance-based classification. Our approach improves upon traditional Euclidean-based methods by incorporating data covariance through support and context prototypes. We specifically focus on scenarios with only just a few annotated images, reflecting real-world limitations where large-scale labeling is not feasible. Validated on pineapple detection from drone imagery, our method outperforms state-of-the-art (SOTA) few-shot models using minimal labeled data. Extensive experiments show that FastSAM, when combined with a Mahalanobis distance variant that applies singular value decomposition (SVD) and diagonal loading for regularization, achieves the highest mean average precision (mAP), offering a practical and effective tool for crop monitoring and management.

Reinforcement Learning has been widely applied to play classic games where the agents learn the rules by playing the game by themselves. Recent works in general Reinforcement Learning use many improvements such as memory replay to boost the results and training time but we have not found research that focuses on the impact of memory replay in agents that play simple classic video games. In this research, we present an analysis of the impact of three different techniques of memory replay in the performance of a Deep Q-Learning model using different levels of difficulty of the Pacman video game. Also, we propose a multi-channel image - a novel way to create input tensors for training the model - inspired by one-hot encoding, and we show in the experiment section that the performance is improved by using this idea. We find that our model is able to learn faster than previous work and is even able to learn how to consistently win on the mediumClassic board after only 3,000 training episodes, previously thought to take much longer.