Innovations in computer vision have steered research towards recognizing compound facial emotions, a complex mix of basic emotions. Despite significant advancements in deep convolutional neural networks improving accuracy, their inherent limitations, such as gradient vanishing/exploding problem, lack of global contextual information, and overfitting issues, may degrade performance or cause misclassification when processing complex emotion features. This study proposes an ensemble method in which three pre-trained models, DenseNet-121, VGG-16, and ResNet-18 are concatenated instead of utilizing individual models. It is a significant layer-sharing method, and we have added dropout layers, fully connected layers, activation functions, and pooling layers to each model after removing their heads before concatenating them. This enables the model to get a chance to learn more before combining the individual learned features. The proposed model uses an early stopping mechanism to prevent it from overfitting and improve performance. The proposed ensemble method surpassed the state-of-the-art (SOTA) with 74.4% and 71.8% accuracy on RAF-DB and CFEE datasets, respectively, offering a new benchmark for real-world compound emotion recognition research.