.png){kind=icon}
.png){kind=icon}
Return
Spikeyolo-RS: Improved Spike-Convolution Neural Network for Remote Sensing Image Object Detection
.png){kind=icon}
PRE
AI
DOI：10.1007/s10812-026-02063-8.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The problem of remote sensing image object detection is one of the important research contents of visual image recognition tasks. Compared with convolutional neural networks, the third-generation brain-like spike neural networks have great potential advantages for on-orbit processing of remote sensing images (RSI) due to its high energy efficiency and low consumption. This paper studies how to increase accuracy of SpikeYOLO for RSI analysis, and proposes an improvement method. First, a dynamic membrane potential attenuation mechanism is constructed and the fixed attenuation factor is reconstructed into a learnable parameter. Then, the membrane potential is updated through vectorized parallel computing. Learnable residual weights are introduced to improve the feature fusion capability and; finally, the calculation fl ow is adjusted. On the RSOD remote sensing dataset, we obtained 96.8% mAP50 and 64.8% mAP50:95, which are 6.7% and 2.3% higher than the previous state-of-the-art SpikeYOLO, respectively. On the NWPU-VHR-10 dataset, we obtained 92.8% mAP50, which is 1.5% higher than SpikeYOLO with the same architecture.
Keywords:
remote sensing image
object detection
spiking neural network
YOLO
SpikeYOLO
Journal
J
IF:
1
Papers:
147
Citations:
1.8K