Research On Remote Sensing Image Target Detection Algorithm Based On Weakly Supervised Collaborative Learning

Remote sensing image object detection is an important task in the direction of computer vision,which has been widely used in real life.At present,the number of remote sensing image labels at the boundingbox-level is small,and the cost of manually labeling remote sensing images is very high.In addition,because the scale of the objects in remote sensing images varies greatly,especially the small-scale objects among them,the feature is not obvious,which makes the small objects in large-scale remote sensing images are easy to miss or misdetect,therefore,improving the detection precision of small objects has also become a key and challenging problem.For the problem of above detection difficulties,this paper proposes effective solutions,the main work contents are as follows:(1)Aiming at the problem that the number of complete strong labels in remote sensing images is small,this paper proposes a collaborative learning algorithm for remote sensing image object detection combining weakly supervised network and strongly supervised network.The whole network only uses rough image-level labels as supervision information for training,which greatly reduces the cost of manual labeling.Moreover,the cooperative enhancement learning of the two sub networks is realized by the consistency loss function,which improves the detection performance effectively.(2)This paper constructs a deep residual network as the feature extraction backbone network,and constructs an enhanced multi-scale fusion network based on this backbone network,which fully integrates low-level location features with high-level semantic features,effectively solves the detection difficulties caused by scale change and weak feature expression capabilities of small objects.(3)This paper proposes an aligned ROI pooling strategy based on bilinear interpolation.The pooling strategy retains decimals during the calculation process,does not perform quantization and rounding,and retains the spatial position of region of interest(Region of Interest,ROI)better,especially improves the detection accuracy of remote sensing small objects.(4)In this paper,three identical fine-tuning network branches are added to the weakly supervised subnetwork,in which the region proposal with the highest detection score is used as the pseudo-labelling of the next branch,and the detection accuracy is improved through continuous adjustment.In addition,the EIo U loss regression function is also defined in RPN of strongly supervised network to obtain more accurate detection results of remote sensing images.