Small Object Detection with Multiscale Features

Guo X Hu,Jia M Han,Li Huang,Zhong Yang,Lei Hu

doi:10.1155/2018/4546896

Abstract

The existing object detection algorithm based on the deep convolution neural network needs to carry out multilevel convolution and pooling operations to the entire image in order to extract a deep semantic features of the image. The detection models can get better results for big object. However, those models fail to detect small objects that have low resolution and are greatly influenced by noise because the features after repeated convolution operations of existing models do not fully represent the essential characteristics of the small objects. In this paper, we can achieve good detection accuracy by extracting the features at different convolution levels of the object and using the multiscale features to detect small objects. For our detection model, we extract the features of the image from their third, fourth, and 5th convolutions, respectively, and then these three scales features are concatenated into a one-dimensional vector. The vector is used to classify objects by classifiers and locate position information of objects by regression of bounding box. Through testing, the detection accuracy of our model for small objects is 11% higher than the state-of-the-art models. In addition, we also used the model to detect aircraft in remote sensing images and achieved good results.

Highlights

Object detection, which requires accurate classification of objects in images and needs accurate location of objects is an automatic image detection process based on statistical and geometric features
The existing detection models based on deep neural network are not able to detect the small objects because the features of objects that are extracted by many convolution and pooling operations are lost
Our model ensures the integrity of the feature of the large object and preserves the full detail feature of the small objects by extracting the multiscale feature of the image

Summary

Introduction

Object detection, which requires accurate classification of objects in images and needs accurate location of objects is an automatic image detection process based on statistical and geometric features. Because different types of images are characterized by different features, it is difficult to use one or more features to represent objects, which do not achieve a robust classification model. These methods (e.g., RCNN [4], Fast-RCNN [5], Faster-RCNN [6], SPP-Net [7], and R-FCN [8]) have achieved good results in multiobject detection in images Most of these object detection algorithms are based on PASCAL VOC dataset [9] for training and testing. The detection model based on the dataset composed of large objects will not be effectively detected for the small objects in reality [10].

Related Works

Model Introduction

Experimental Analysis

Findings

Conclusions