Research On Robust Algorithm Suitable For Small Object Detection

Because of the characteristics of small object with small size and low resolution,it has always been a difficult problem for object detection.It is necessary to design a suitable detection model according to the characteristics of small object.From the perspective of feature enhancement,this paper proposes novel feature fusion module,attention module and feature pyramid module.These modules are applied to the classical detection model to improve the accuracy of small object detection.The following are two main tasks of this paper to improve the detection accuracy of small object:1)The first work is to design a novel attentional module to improve the representation capability of feature maps.This module can capture global semantic information while modeling the relationship between channels,which makes up for the shortcoming that convolution operation can only capture local features.With the help of global semantic information,the recognition effect of small object is improved in combination with its environment.Various experiments were designed on PASCAL VOC dataset and RSOD remote sensing satellite dataset to verify the enhancement effect and generalization ability of this module on different detection models.2)The second work is to improve SSD object detection algorithm to enhance its ability to identify small object.In order to further improve the performance of the SSD(Single Shot Multibox Detector)algorithm and solve the problem of the imbalance of feature map information in the multi-scale prediction of the SSD algorithm,a plug-and-play module is designed in this paper to enhance the representation ability of feature map.Firstly,a novel feature fusion method is designed to solve the problem of information disparity in cross-scale feature fusion.Secondly,according to the idea of pooling pyramid,a depth feature extraction module is designed to extract the information of different receptive fields,so as to improve the ability of object detection.Finally,the architecture of SSD algorithm is improved by using the proposed module.Through the above mechanism,the detection accuracy and robustness of SSD algorithm are effectively improved.Extensive experiments have been conducted on PASCAL VOC datasets to verify the efficiency of the proposed method.On Pascal VOC2007 test datasets,our method improves 2.9% over SSD algorithm,while maintaining the ability of real-time detection.3)The third work is to improve the FPN module to further enhance the information fusion ability of feature pyramid and improve the accuracy of small object detection.The top-down fusion mode of FPN results in that the rich information of shallow feature maps cannot be effectively transferred to deep feature maps,resulting in the information imbalance of feature maps.In addition,FPN directly fuses two feature maps with different information easily to introduce noise information,which is not conducive to small object detection.In order to solve these two problems,DFFPN Network(Deep Fusion Feature Pyramid Network)is proposed in this paper.Conditional convolution is used to extract the common information of the two Feature maps as the template,and subsequent Fusion based on the template Feature maps can effectively reduce the introduction of noise information.In order to enhance the ability of information transfer,cross-residual link is adopted to overcome the one-way information flow problem of top-down fusion.The experimental results show that the detection accuracy of small object is improved to varying degrees,among which the Faster-RCNN model is improved by2.4% and models based on Retina Net improved by 2.3% in the RSOD remote sensing satellite dataset.