Improved Image Classification using Lightweight Deep Neural Network Enhancements

In this paper, a novel hierarchical deep neural network (DNN) is introduced that augments an input DNN to significantly enhance its image classification accuracy while reducing the inference time and the hardware overhead. The architecture comprises a hybrid framework that combines binary classifiers based on Convolutional Neural Networks (CNNs) with refined classifiers employing Vision Transformers (ViTs). A distinctive training approach is employed, where embedded models are designed and trained based on image distributions processed by binary classifiers, enhancing the system's precision and efficiency. An algorithm determines the optimal inclusion of components within a cascading structure, enabling the construction, training, and deployment of specialized deep-learning networks. Additionally, two algorithms are introduced to optimize the architecture for multi-GPU systems. Extensive experimentation across multiple baseline DNNs, including both CNNs and ViTs, and diverse datasets demonstrates the versatility and superiority of our proposed structure over traditional methods, with particularly strong improvements observed on larger and more complex datasets such as ImageNet.