1. Reference article: "Remote Sensing Target Detection and Recognition convolutional neural network-based"
Summary:
The problem is the conventional training method based on Remote Sensing Target convolutional neural network dependent on large bounding box data (position information data), It requires a lot of manual annotation cost, and a limited number of image sensing target sample, not enough to support large-scale training; additional conventional image sensing target recognition method based on convolutional neural network considering only deep semantic characteristics of the network , resulting in recognition performance reached a bottleneck.
In this paper, CNN feature extraction region extracted by the depth of interest, and by a plurality of scales CNN confirmed the target region of interest, without bounding box training data.
Contribution points:
(1) to solve the target detection methods rely on a large number of bounding box data and insufficient training of remote sensing images of the target sample problem. Resolution Images trained using a convolutional neural network, using the network characteristics of extracting depth resolution remote sensing image acquisition area of interest, the final multi-scale neural network convolution target confirmation. Benefits: improve the accuracy of target detection, reduce the missing rate, robust, bounding box data without the need to train significantly reduces training time.
(2) First, a convolutional neural network is designed for remote sensing image more target recognition, then extracting the features of one end of the convolution into the network, the network features extracted by the different depths of average global pooling method, its linear integration.
Chapter 1 Introduction
1.1 Background and Significance of
the previous algorithm mainly sliding window algorithm, a method based on image analysis, object recognition algorithm BOW feature-based object recognition algorithm based on HOG features.
Probably, these features can not express the abstract goal of semantic features. What is abstract semantic feature? ? CNN can be well detected and classified because CNN shared weights, the displacement rotational invariance. Deep semantic features extracted CNN can be effectively describes the differences between the different categories of natural scene image. Edge, texture, color is superficial characteristics; abstract high-level semantic features (features include abstract semantics, semantics can effectively distinguish between different targets, such as airports, ports), meaning that high-level abstract semantic features can be categorized it.
. . .
Chapter 2 related work
2.1 and Remote Sensing Target Recognition
Currently, remote sensing images based on machine learning target detection algorithm is divided into two parts, the region of interest and a target confirmation section extraction section, wherein the target confirmation section mainly Some machine learning models extracting an image feature, and train a classifier using a classifier to classify regions of interest confirmed. . . .
Chapter 3,
finally, finally, seems to get to know the ideas of this paper. Depth features seemed to him quite cattle, introduced a lot of classical methods, and finally leads to convolution neural network is proposed target detection method for remote sensing image based on the depth of features. But in the current knowledge, no features are based on the depth of it? ?
Faster R-CNN and YOLO need a lot of training data for large-scale training, including category tag data and bounding box data. Remote Sensing Target does not have the large-scale training data set. . .
Accordingly, we proposed a new method for remote sensing images based on the depth of the target detection feature. First, a high-resolution 512 * 512. . . .
2. Reference article: "Based on Remote Sensing Target convolution neural network"
Abstract: In
order to solve the remote sensing images of the target rotation angle, will transform the network into the space Faster R-CNN, proposes a target rotation invariant self-learning ability detection model. Melting space transformation network convolutional neural network having better characteristics referred to achieve rotational invariance.
Introduction:
Remote Sensing Target presents a top view perspective, a large angle range and direction. Therefore, the features extracted object having the rotational robustness.
CNN senior feature extraction, compared to manual extraction of more abstract features, has better robustness. Thanks to pooling layer, convolution neural network has some ability to learn to pan and zoom, but still lack the ability to learn rotational invariance of the input data. STN is a transform module. By displaying learning, traditional convolutional layer obtained can pan, zoom, rotate, and other characteristics. This article Faster R-CNN and STN module proposed by combining the rotation of the target detection more robust framework. . . .
3.参考文章：《OBJECT DETECTION IN SATELLITE IMAGERY USING 2-STEP CONVOLUTIONAL NEURAL NETWORKS》
Abstract
efficent
proposed method: two convolution neural networks
effects: higher accuracy (recall, precision, F-measure: 87%)
Introduction
High resolution satellite imagery can detect small objects such as ships, cars, aircraft, and individual houses; whereas, medium resolution satellite imagery can detect relatively larger objects, such as ports, roads, airports and large buildings [1] [6]
As an example of target object, we selected golf courses because they exist everywhere in the world, are typically of a recognizable size and shape with 30 meter resolution of Landsat 8 imagery data sets described:. Landsat 8 imagery of the golf course
if Landsat8 datasets pictures can be selected according to the year, area.
4.参考文章：《You Only Look Twice: Rapid Multi-Scale Object Detection In Satellite Imagery》
Abstract
solve the problem: a pixel, a wide range of small targets of interest
to provide a method: YOLT
effect: at a rate of 0.5km2 / s speed image sensing evaluate any dimension
rapid detection of a variety of scales the image
F1> 0.8
, even if the pixel size is only 5 high confidence target can still be detected.
Introduction
to Remote Sensing Target task, mainly in the following issues:
- small spatial extent: the target is very small and dense
- complete rotation invariance: almost any rotation
- training example frequency: the lack of training data
- ultra high resolution: sampling is not the right choice simple
for the above problem, in turn propose solutions. In order to solve small and dense target, 3.1 proposed a new network architecture. . . .
Training set
training set from three places: DigitalGlobe SATELLITES, Planet SATELLITES, Aerial Platforms.
Labels made a bounding box and the composition of the class. Let's focus on five categories: airplanes, boats, building footprints, cars, airports. For different sizes and objectives, we used two different scales detector is very effective.
Cars: COWC data set is a large, high-quality multi-scale car marked a good data set. Data set from aerial platform
5.参考文章：《REDUCED FOCAL LOSS: 1st PLACE SOLUTION TO XVIEW OBJECT DETECTION IN SATELLITE IMAGERY》——2019.04.29
Abstract
solve the problem: To solve the imbalance problem DIUx xView 2018 Detection Challenge sample data set
proposed method: Reduced Focal Loss function,
the effect of: obtaining first place in this competition. Evaluation criteria: still mAP (31.74), Recall (61.2 %), mRecall (77.5%)
introduction
describes the characteristics of data sets, DIUx xView 2018 Detection Challenge is the largest open Remote Sensing Target data set. It contains about one million target, 60 class. Target large scale will not change, but can be rotated 360 degrees. Difficult to identify data sets of features, and characteristics inherent imbalance.
Methods
related work
each picture have been a backbone network, such as resnet, each block Take a feature map. RPN for each feature map, but all share the same layer weight, thereby generating proposals. The best proposals to use two fully connected network to correct, the FPN first collected from each layer prediction, run NMS. Layer then choose the best use ROIAlign, based on the size of the Proposal. Results detector base FPN Table 1.
Experimental
xView dataset contains 864 labels a good picture, 742 are divided into a training set and a validation set 104.
Training set is cropped to 700 * 700 small images, each small picture data enhancement operation: Rotate, Flip
code implemented pytorch. Of course, there are many details to achieve.
6.参考文章：《Object Detection in Aerial Images Using Feature Fusion Deep Networks》(2019)
Abstract
Since remote sensing images of high density, small size of the target, and the complexity of the scene in the conventional method of remote sensing image accuracy is low. We propose FFDN, can achieve higher detection performance. And detecting in UAV123 UAVDT dataset.
Experimental
data sets
UAV123 data sets of various kinds, UAVDT complex and diverse data sets scene. UAV123 dataset contains 123 videos, we selected 33 videos to produce 48,770 sets of data may comprise a variety of scenes. We produce 13,871, generated manually ground truth. The main object is detected type bikes, boats, buildings, people, cars and the like.
UAVDT data sets are taken at six different urban areas, defined six kinds of attributes.
Evaluation criteria
we use four standard: precision §, recall ®, F1 -score (F1) and mean intersection over union (Mean IoU))
results
7. Reference article: "Clustered Object Detection in Aerial Images" (2019)
Abstract
Remote Sensing Target is difficult, mainly due to the following: . .
Thus ClusDet made
on three test data sets: VisDrone, UAVDT, DOTA
8.参考文章：《Object Detection in Very High-Resolution Aerial Images Using One-Stage Densely Connected Feature Pyramid Network》（2018）
Abstract
Remote sensing images scale and target different appearances, leading to Remote Sensing Target task particularly difficult. So our proposed method can deal with the problem of a different scale. On two datasets disclosed experiments have proven mAP better and less computation time.
Experimental
data set description
NWPU VHR-10 data set, the data set annotation 650 provides good picture, each picture comprising at least one target. These images are also labeled by hand bounding box as ground-truth.
RSOD dataset
evaluation standard
precision-recall curves
AP
results
using three kinds backbone, are VGG-16, Resnet50, Resnet101.
This feeling is better understood, although there is no code.
9.参考文章：《DOTA: A Large-Scale Dataset for Object Detection in Aerial Images》（2018）
10.参考文章：《Salience Biased Loss for Object Detection in Aerial Images》（2018）
. . .
. . .
Data collection for the DOTA
11.参考文章：《Multi-Scale Image Block-Level F-CNN for Remote Sensing Images Object Detection》（2019）
摘要
在遥感图像目标检测领域有很多困难，比如complex and varied appearances, the expensive manual annotation,(人工标注困难)and difficult in fast detecting the large scene image。于是我们提出MIF-CNN。它有特别多的好处。。。。
数据集：NWPU VHR-10，两个Airports数据集。
12.参考文章：《A Sample Update-Based Convolutional Neural Network Framework for Object Detection in Large-Area Remote Sensing Images》（2019）
13.参考文章：《Detection of Multiclass Objects in Optical Remote Sensing Images》（2019）
14.参考文章：《Multiscale Visual Attention Networks for Object Detection in VHR Remote Sensing Images》（2019）
15.参考文章：《Scale Adaptive Proposal Network for Object Detection in Remote Sensing Images》（2019）
16.参考文章：《Object Detection in Remote Sensing Images Based on a Scene-Contextual Feature Pyramid Network》（2019）
17.参考文章：《A Novel Multi-Model Decision Fusion Network for Object Detection in Remote Sensing Images》（2019）
18.参考文章：《A Training-free, One-shot Detection Framework For Geospatial Objects In Remote Sensing Images》（2019）
19. Reference article: "HIERARCHICAL REGION BASED CONVOLUTION NEURAL NETWORK FOR MULTISCALE OBJECT DETECTION IN REMOTE SENSING IMAGES": Northern, Germany
Abstract
We propose a multi-scale objective Faster R-CNN-based method to detect remote sensing image. First, a pre-trained CNN for extracting features from the input picture, and then produce a series of candidates.
Dataset: Google Earth, GaoFen-2
20.参考文章：《Deep Adaptive Proposal Network for Object Detection in Optical Remote Sensing Images 》
Summary
. . .
. . .
Data Set: NWPU VHR-10