TY - GEN
T1 - FCOS Small Target Detection Algorithm Combined with Multi-Layer Hybrid Attention Mechanism
AU - Liu, Ying
AU - Geng, Luyao
AU - Yu, Hao
AU - Xu, Zhijie
N1 - Publisher Copyright:
© 2021 ACM.
PY - 2021/9/24
Y1 - 2021/9/24
N2 - The current target detection algorithm can be competent for most of the detection tasks. However, improving the detection accuracy of small targets is difficult due to the small target occupy less pixels and the feature extraction is hard to achieve. To address this problem, the per-pixel target detection algorithm FCOS is adapted in this research, and the widely used ResNet50 is implemented as the algorithm backbone, by adjusting the size of the input image and the composition of the loss function. The CBAM hybrid attention mechanism is applied into the shallow features and high-level features corresponding to the bottom pyramid, then the feature pyramid is constructed to achieve the purpose of multi-scale target detection. The comparison and ablation experiments show that the original FCOS target detection model can improve the detection accuracy by about 3.7% and the small target detection accuracy by about 2% on the MS-COCO dataset.
AB - The current target detection algorithm can be competent for most of the detection tasks. However, improving the detection accuracy of small targets is difficult due to the small target occupy less pixels and the feature extraction is hard to achieve. To address this problem, the per-pixel target detection algorithm FCOS is adapted in this research, and the widely used ResNet50 is implemented as the algorithm backbone, by adjusting the size of the input image and the composition of the loss function. The CBAM hybrid attention mechanism is applied into the shallow features and high-level features corresponding to the bottom pyramid, then the feature pyramid is constructed to achieve the purpose of multi-scale target detection. The comparison and ablation experiments show that the original FCOS target detection model can improve the detection accuracy by about 3.7% and the small target detection accuracy by about 2% on the MS-COCO dataset.
KW - Anchor-Free
KW - Convolutional neural network
KW - Hybrid Attention Mechanism
KW - Object detection
UR - http://www.scopus.com/inward/record.url?scp=85125836912&partnerID=8YFLogxK
U2 - 10.1145/3488933.3489000
DO - 10.1145/3488933.3489000
M3 - Conference Proceeding
AN - SCOPUS:85125836912
T3 - ACM International Conference Proceeding Series
SP - 50
EP - 55
BT - AIPR 2021 - 2021 4th International Conference on Artificial Intelligence and Pattern Recognition
PB - Association for Computing Machinery
T2 - 4th International Conference on Artificial Intelligence and Pattern Recognition, AIPR 2021
Y2 - 17 September 2021 through 19 September 2021
ER -