Balance Loss for MultiAttention-Based YOLOv4
Zijing Niu, Bo Jiang, Hong Mei Xu, Yuhang Zhang · 2023
Aiming to address the advantages and disadvantages of various target detection algorithms, we optimized the data augmentation algorithm, attention mechanism and loss backpropagation function across different neural networks, and proposed an improved YOLOv4-based target detection algorithm. The new model introduces a Mixup algorithm with Mosaic in the data augmentation block, adds a multi-attention block that includes channel attention and spatial attention between the backbone and FPN, and replaces CIoU in the Loss function with Balance IoU and cascade confidence loss called Balance Loss to enhance the neural network's target detection capability. We trained and validated the improved algorithm in the VOC2007 dataset, and experimental results showed that this algorithm has a more robust feature extraction capability and higher target detection accuracy than the original YOLOv4 algorithm. The mAP increases by 1.2%. The model also performs better in detecting small and medium target objects.