Dual-Path Autoencoder-DenseNet Architecture for Robust Object Classification via Background-Free Reconstruction and Multi-Task Feature Fusion

Object classification in real-world environments is often challenged by image degradation such as noise, blur, and poor lighting, which degrades the performance of conventional deep learning models. To address this, we propose a multi-task dual-path autoencoder-DenseNet framework that simultaneously performs image reconstruction and object classification within a unified architecture. The model utilizes a dual-path design, incorporating a DenseNet121 backbone to extract multi-scale features for both image reconstruction and classification tasks. By using a shared encoder, the model extracts deep features from degraded inputs, reconstructs a clean foreground-focused image, and classifies the object in a seamless pipeline. Our approach employs a diverse set of loss functions, including Mean Square Error, Root Mean Square Error, Mean Absolute Error, and Huber, to optimize both tasks. Experimental results on an augmented Edinburgh Kitchen Utensils Database (EKUD) show the best performance with $\mathbf{8 1. 0 0 \%}$ classification accuracy on Mean Square Error reconstruction loss function, outpacing conventional models and sequential reconstruction-classification pipelines. This multi-task dual-path architecture offers a robust, end-to-end solution for real-world visual challenges, with applications in autonomous driving, industrial inspection, and robotics.