Lightweight faster R-CNN for object detection in optical remote sensing images.

Various applications in remote sensing rely on object detection approaches, such as urban detection, precision farming, and disaster prediction. Faster RCNN has gained popularity for its performance but comes with significant computational and storage demands. Model compression techniques like pruning and quantization are frequently employed to mitigate these challenges. This paper introduces a novel bi-stage compression approach to create a lightweight Faster R-CNN for satellite images with minimal performance degradation. The proposed approach employs two distinct phases: aware training and post-training compression. First, aware training employs mixed-precision FP16 computation, which enhances training speed by a factor of 1.5 to 5.5 while preserving model accuracy and optimizing memory efficiency. Second, post-training compression applies unstructured weight pruning to eliminate redundant parameters, followed by dynamic quantization to reduce precision, thereby minimizing the model size at runtime and computational load. The proposed approach was assessed on the NWPU VHR-10 and Ship datasets. The results demonstrate an average 25.6% reduction in model size and a 56.6% reduction in parameters while maintaining the mean Average Precision (mAP).