1、Feature-Fused SSD: Fast Detection for Small Objects (ICGIP2017)
Fusion features SSD
[1709.05054] Feature-Fused SSD: Fast Detection for Small Objects https://arxiv.org/abs/1709.05054
Paper: https://arxiv.org/ftp/arxiv/papers/1709/1709.05054.pdf
Rapid detection of small targets, the SSD as a basic structure, for both speed and accuracy.
The idea is to use high-level semantic features to enhance FIG low-level feature maps in order to increase the accuracy of detection of small objects.
Wherein the different layers to obtain fusion contextual information (context information) of the method, and apply it to the model SSD.
SSD using feature detection pyramid in order to take into account the objectives of different scales, the most superficial features used on small target detection, because of the shallow feel of a small field, just receptive field size and small targets match.
Characterized in shallow lack semantic information (Semantic information). Semantic information can affect the detector determines the target detection region (object) or a background (background). So will feature high-level and low-level feature fusion, will be receptive fields fit, but not a lack of features semantic information.
Element-Sum module concatenation module and the two fusion ways: using the splicing, concat or sum operation.
concat enation Module can reduce the effects of occlusion and element- Sum Module for better results and less fuzzy pixels of the target.
【forward from】
Rapid target detection --Feature-Fused SSD: Fast Detection for Small Objects - CSDN blog - AI small workshops blog https://blog.csdn.net/zhangjunhit/article/details/78031452
Small Target Detection: Feature-Fused SSD - starfish blog - CSDN blog https://blog.csdn.net/weixin_39290638/article/details/84540806
Feature-Fused SSD: Fast Detection for Small Objects - DCD_LIN 's blog - CSDN blog https://blog.csdn.net/linmingan/article/details/81095435
Fusion features of SSD: for rapid detection of small targets - blog wangxujin666 of - CSDN blog https://blog.csdn.net/wangxujin666/article/details/83065261
2、FPN（feature pyramid networks）
Thesis: the Feature Pyramid Networks for Object Detection 5 FPS
papers link: https://arxiv.org/abs/1612.03144
This is cvpr2017年出的一篇文章, this article is mainly for detection of common goals, but played a key role in small target detection, many small target detection method after that have used a similar method, such as faster-rcnn + fpn, yolov3 the feature fusion.
| In order to achieve higher positioning accuracy while ensuring the speed, yolo v3 using a more complex network structure. Compared to the previous network, the improvement comprising yolo v3 multiscale prediction (the FPN), the structure of more complex network Darknet53 canceled softmax classified as a candidate block, which are made yolo v3 faster speed, accuracy rate corresponding improve. |
Features mode using the multi-scale fusion, using features result after fusion Different layers do prediction.
Low-level features of semantic information is relatively small, but the target position accuracy, high-level semantic feature rich, but a rough position of the target.
After do prediction using a top-down process after FPN, wherein each layer obtained in FIG method on small features after the sampling FIG fused FIG next feature, fusion, thus sequentially, to obtain a plurality of predictors .
Feature Pyramid Network
Algorithm idea: authors propose a path from top to bottom and lateral path. It refers to the path from top to bottom of the deep network structure wherein FIG upsampling operation, it is consistent with the size of a network of shallow wherein FIG, wherein FIG stacking can be performed. It refers to the final lateral path characteristic diagram for each of the convolution modules using 1X1 convolution kernel down-channel operation, reduce the number of features of FIG. Stacked end-use multi-scale characteristic diagrams model learning classification and positioning.
Algorithms effect: the feasibility of FPN practices, the authors note, shallow semantic features characteristic map is weak, and the strong position of features; rather, semantic features deep drawing characteristics of a strong, but the loss of precise location features. While the use of shallow and deep FIGS features characteristic graph can be characterized into consideration the location and intensity strong semantic features, thus to enhance the effect of the model.
Fully Convoluntional Network
Algorithms thought: like the whole idea of the FPN. The innovation is that (1) a full connection layer discarded, replaced whole connection layer equivalent 1X1 convolution kernel, so that the input image can be different scale networks. (2) characteristics of the stack of FIG sampling continues on, so that its original size and consistency. Stacking features of FIG upsampled, classification prediction made at the point where it is mapped to the original pixel position. In this way, artwork may be made based on a fine image segmentation.
Algorithms effect: FCN create fine segmentation of its kind, MASK-RCNN after its algorithms also draws its ideology. For target detection, classification by pixel, to make finer division position.
【forward from】
Small Target Detection article summary - wq604887956's blog - CSDN blog https://blog.csdn.net/wq604887956/article/details/83053927
Small Target Detection paper reading - happy dragon fruit blog - CSDN blog https://blog.csdn.net/weixin_40683960/article/details/82078957
【other】
https://cloud.tencent.com/developer/article/1414911
Small Target Detection: Improving Small Object Detection - starfish blog - CSDN blog https://blog.csdn.net/weixin_39290638/article/details/84487984
Convolution neural network can be used to detect small targets it? - know almost https://www.zhihu.com/question/49722539