Modified Asymmetric Focal Loss for Fairness Improvement in Object Classification

ABSTRACT Class imbalance is a well‐recognized challenge in object classification, where dominant classes often bias the learning process, resulting in reduced accuracy and fairness issues for underrepresented categories. This problem is particularly critical in aerial imagery datasets, where the distribution of air‐to‐ground objects is highly skewed. To address this issue, the present work systematically investigates the role of loss functions in mitigating class imbalance and proposes novel variants of asymmetric focal loss (MAFL) that enhance balanced performance across categories. This proposed MAFL approach introduces adaptive modulation through balancing and focusing parameters to prevent class neglect and overfitting to rare instances. By integrating fairness awareness within the loss function, it ensures balanced learning, enhanced detection accuracy and equitable class representation. Extensive experimentation has been conducted on the VisDrone and UAVDT datasets, which provides a diverse and challenging benchmark for air‐to‐ground object classification. The proposed variants consistently improve the recognition of minority classes while maintaining strong overall accuracy, demonstrating their effectiveness in handling skewed distributions in aerial data. In addition to empirical validation, this study introduces a benchmarking framework that enables structured comparison of loss functions under imbalanced settings, offering a reproducible reference for future research in fairness‐aware classification. The source code for this loss function is publicly available at: https://github.com/battishneeraj/FocalLoss_Code .