Deep learning methods have exhibited the great capacity to process object detection tasks, offering a practical and viable approach in many applications. When researchers have advanced deep learning models to improve their performance, the model derived from the algorithmic improvement may itself require complementary increases in computational and power demands. Recently, model compression and pruning techniques have received more attention to promote the wide employment of the DNN model. Although these techniques have achieved a remarkable performance, the class imbalance issue during the mode compression process does not vanish. This paper exploits the Autonomous Binarized Focal Loss Enhanced Model Compression (ABFLMC) model to address the issue. Additionally, our proposed ABFLMC can automatically receive the dynamic difficulty term during the training process to improve performance and reduce complexity. A novel hardware architecture is proposed to accelerate inference. Our experimental results show that the ABFLMC can achieve higher accuracy, faster speed, and smaller model size.
Keywords: embedded hardware; focal loss; tensor decomposition.