Multi-Scale Object Detection With Convulution Neural Network

In this paper,the multi-scale object detection problem is studied.We explore leveraging multi-scale fusion feature of deep convolutional neural networks to promote the performance of multi-scale object detection,especially the ability of small object detection.In this paper,firstly,the task of airplane detection in high resolution optical remote sensing imagery is taken as an example.Based on the SSD(Single Shot MultiBox Detector)detection algorithm using multi-scale representation of CNN(Convolutional Neural Nerworks),the multi-scale object detection problem under large-scale and highresolution remote sensing images is studied.This paper describes the phenomenon of "semantic gap" between different scales of convolutional neural networks and proposes a data-driven hyperparameter selection method to make the multi-layer detection layer more balanced and fully trained.In addition,multi-scale sample training is also used to improve the performance of lower-level detectors.Based on the CNN multi-scale SSD object detection algorithm,this paper investigates the fusion method of multi-scale features,and proposes an improved SSD object detection algorithm based on multi-scale feature fusion to promote the semantic discrimination of each scale feature and solve the problem of different scale features.In order to overcome the sample class imbalance of single-step detection algorithm,this paper proposes a multi-scale target detection algorithm based on multi-scale feature fusion SSD algorithm,which combines multi-scale fusion features and object priori.By leveraging the object priori,the SSD detection algorithm can be controlled to update the network weights only in the area where the suspected object exists,thus significantly reducing the object search space and alleviating the imbalance of positive and negative samples.In this paper,we train and test our improved algorithm on the optical remote sensing image dataset and the KITTI vehicle test dataset.Experiments show that the final proposed detection algorithm combined with multi-scale features and object prior has reached state of the art.