Efficient objectdetection from optical remote sensing (RS) images has always been an important interpretation task for in-orbit RS applications. In recent years, convolutional neural networks have been widely used fo...
详细信息
ISBN:
(数字)9789819715688
ISBN:
(纸本)9789819715671;9789819715688
Efficient objectdetection from optical remote sensing (RS) images has always been an important interpretation task for in-orbit RS applications. In recent years, convolutional neural networks have been widely used for objectdetection with significantly improved detection accuracy. However, the large detection models pose great challenges for the computing, memory and energy supply of resource-constrained in-orbit platforms. In this paper, we propose an efficient in-orbit object detection method with low memory, computation and energy requirements. The proposed method first integrates the compact modules of GhostNet into the detector and further performs the L1-norm based filter pruning to significantly reduce model size and computational complexity. Besides, we propose to use energy as a key metric in filter pruning, and present a novel energy-guided layer-wise pruning rate estimation method so as to achieve energy-efficient objectdetection. Comprehensive experiments have shown the effectiveness of the proposed method in terms of model size, computational complexity, latency and energy consumption, while maintaining comparable detection accuracy.
We are studying in-orbit real-time objectdetection for remote sensing satellites. Due to the small object size of remote sensing images, it is hard to achieve high detection accuracy, especially for resource-constrai...
详细信息
ISBN:
(纸本)9781510655386;9781510655379
We are studying in-orbit real-time objectdetection for remote sensing satellites. Due to the small object size of remote sensing images, it is hard to achieve high detection accuracy, especially for resource-constrained spacecraft computers. Lightweight objectdetection models such as YOLO and SSD are feasible choices to achieve acceptable detection speed on board. This study proposes an accuracy-improvement method for the lightweight neural networks with an upscaling ratio estimator without retraining the model. The estimator exploits a scaling ratio that determines how much the image should be resized. With our scaling estimator, we have achieved 10.09% higher accuracy than the original YOLOv4-Tiny models with a 40% detection speed overhead.
暂无评论