<p>Conventional deep learning image analysis methods typically aim at a single task, e.g., classification or reconstruction, and are unable to generalize to a wide range of visual tasks. This disparity inhibits the emergence of models that can learn image representations that are comprehensive and that can be transferred across. The objective of the study is to design a powerful multitask deep learning architecture that not only carries out image classification but it also reconstructs to improve the learning of features as well as predictive rates. A hybrid supervised unsupervised learning model has been developed as encoder-decoder architecture with a classification head. The model used was trained and tested on the CIFAR-10 dataset of 60000 labelled images in ten classes. Preprocessing and augmentation of data were done to enhance sample diversity. The loss function was a mixed loss having cross-entropy loss (weight: 1) as the classification loss and mean squared error (weight: 0.1) as the reconstruction loss. To achieve a smooth convergence of the network, the learning rate set was 0.001 and the dropout rate was 0.3. The proposed model had a classification accuracy of 89, precision, recall, and F1-score of 88.8, 88.5 and 88.6 respectively. Reconstruction error was 0.025, which showed that the reconstruction quality was good. Computational efficiency was established by having the overall training time of about 700&#xa0;s. The performance confirms the usefulness of the suggested multitask learning method when it comes to strike a balance between classification and image reconstruction. The model is a scaled, interpretable, and generalizable answer that can be applied to multitask deep learning applications in computer vision.</p>