We found a match
Your institution may have access to this item. Find your institution then sign in to continue.
- Title
CALYOLOv4: lightweight YOLOv4 target detection based on coordinated attention.
- Authors
Wang, Huilin; Qian, Huaming; Feng, Shuai; Yan, Shuya
- Abstract
The current deep learning-based target detection algorithm YOLOv4 has a large number of redundant convolutional computations, resulting in much consumption of memory and computational resources, making it difficult to apply on mobile devices with limited computational power and storage resources. We propose a lightweight YOLOv4 (CALYOLOv4) target detection algorithm based on coordinated attention to solve this problem. First, we use MobileNetv2CA with a coordinated attention mechanism instead of CSPDarknet53 as the backbone feature extraction network to reduce network parameters and improve network attention. Second, we use depthwise separable convolutions and mixed depth convolutions (MixConv) to replace the standard convolution in the network, further reducing the parameters and computation of the network. Finally, we choose a better-weighted bidirectional feature pyramid (BiFPN) to replace PANet as the feature fusion network to fully fuse features between different scales. The test results on the PASCAL VOC and MS COCO datasets show that, compared with the YOLOv4 algorithm, our proposed CALYOLOv4 algorithm has 89.1% fewer model total parameters and is 1.71 times faster, reaching 65 frames per second on NVIDIA GeForce RTX 3060, with 81.0% and 29.6% detection accuracy, respectively, achieving the best balance of accuracy and speed. The feasibility and effectiveness of the proposed algorithm are fully demonstrated.
- Subjects
DEEP learning; MOBILE learning; FEATURE extraction; POWER resources; PROBLEM solving
- Publication
Journal of Supercomputing, 2023, Vol 79, Issue 16, p18947
- ISSN
0920-8542
- Publication type
Article
- DOI
10.1007/s11227-023-05380-3