Method and implementation of region proposals extraction for complex scenes

Land remote sensing images are large in width, low in effective object density and complex in distribution. Using the existing detection strategies to finely process each pixel of the remote sensing images is difficult to meet the detection needs of high accuracy and fast speed for some tasks. It is necessary to extract regions proposals effectively. To this end, this paper designs a classification network based on embedded linear expansion features to achieve fast and accurate screening of region proposals. Specifically, we first optimally design the embedded linear expansion feature module to achieve lightweight convolution; secondly, we design the residual structure of the embedded linear expansion feature through the stacking module, to effectively extract feature; finally, we construct a classification network based on the embedded linear expansion feature through the combination of the proposed residual structure, which can classify the sliced image quickly and extract the region proposals effectively. The test results on the public dataset DOTA show that the proposed method with suitable detection has a significant improvement in accuracy and speed compared to the baseline model YOLOV4, with a 4% improvement in mAP and a reduction in computation to 16.8 times of the original.