Improving Tiny YOLO with Fewer Model Parameters

Yanwei Liu, Ching-Wen Ma

Abstract: With the rapid development of convolutional neural networks (CNNs), there are a variety of techniques that can improve existing CNN models, including attention mechanisms, activation functions, and data augmentation. However, integrating these techniques can lead to a significant increase in the number of parameters and FLOPs. Here, we integrated Efficient Channel Attention Net(ECA-Net), Mish activation function, All Convolutional Net (ALL-CNN), and a twin detection head architecture into YOLOv4-tiny, yielding an AP <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">50</inf> of 44.2% on the MS COCO 2017 dataset. The proposed Attention ALL-CNN Twin Head YOLO (A <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> -YOLO) outperforms the original YOLOv4-tiny on the same dataset by 3.3% and reduces the model parameters by 7.26%. Source code is at https://github.com/e96031413/AA-YOLO