Attention Aware Dual-Path Autoencoder with Asymmetric Loss for Recognition in Complex Scenes

Object recognition in complex scenes is challenging due to cluttered backgrounds, overlapping objects, and degraded image quality. Another difficulty arises from sparse label presence, as most images contain only one to three active labels despite the dataset being balanced across 20 object classes. This intra-sample sparsity complicates binary classification by exposing models to a high proportion of inactive classes. This work aims to improve recognition accuracy, robustness under sparse multi-label conditions, and interpretability in visually complex environments. The objective is to help models focus on relevant visual features, suppress background noise, and better distinguish objects that are rare or overlapping. To address these challenges, we introduce an attention aware dual-path autoencoder that enhances image features while learning to classify multiple objects. The model uses asymmetric loss to reduce the influence of easy negatives and emphasize rare or difficult labels. It also integrates an attention mechanism in the reconstruction path to improve object clarity. The proposed model achieves 96.72 percent accuracy, 0.0328 Hamming Loss, 0.9809 macro ROC-AUC, and 0.8925 macro mAP, along with 0.9372 SSIM and 7.1012 dB PSNR in reconstruction. These results confirm its effectiveness for robust classification and enhanced visual understanding in complex scenes.