How to realize fast recognition of oversized images

 In general , Remote sensing target detection , The image size of remote sensing image will be very large , And the elements in the image are extremely complex , Amazon cloud technology, which recently started the race 【AI For Good - 2022  Remote sensing optical image target detection challenge 】 No exception , Easily exceed 10000 x 10000 The satellite remote sensing images made many players feel headache . At the same time, the target size in remote sensing images varies greatly 、 Small and dense 、 Different angles also lead to common CV The framework is difficult to achieve fast and accurate target recognition . therefore , How to realize fast recognition of large-scale images such as remote sensing images ？

Copy

At present, there are many mature satellite image recognition algorithms , But most of them rely on powerful computing resources , In order to realize large-scale image recognition with limited computing resources , We found a feasible open source framework , It provides a good idea for large-scale image recognition .

YOLT  It's based on YOLO v2 Open source algorithm for satellite image recognition , The core idea is ：

1. Solve the problem of image size through image cutting and image network reconstruction ;2. adopt “ On the sampling ” Improve the detection accuracy of small and clustered targets ;3. By fusing different size models , Improve the overall detection accuracy .

YOLT Project address ：GitHub - avanetten/yolt: You Only Look Twice: Rapid Multi-Scale Object Detection In Satellite Imagery

YOLO It is a classic image recognition algorithm ,YOLT stay YOLO On the basis of satellite image, this paper puts forward specific solutions to the unique problems of satellite image . See the paper for detailed ideas 「You Only Look Twice: Rapid Multi-Scale Object Detection In Satellite Imagery」（GitHub There is a link to the paper in the project ）.

▲ On the left is the frequently asked questions , On the right is the solution

1. For objects Irregular size 、 Diverse directions The problem of ,YOLT Data enhancement processing such as size transformation and rotation of satellite image data .

2. in the light of The target size is too small and clustered The problem of ,YOLT The framework mainly adopts 3 There are ways to deal with ：

(1)  Modify the image network structure , take YOLO v2 In the framework stride from 32 Change it to 16, It is helpful to detect that the size is 32 x 32 The following goals

(2)  Upsampling the image , Complete the of the picture “ decompression ” operation , That is to enlarge the original picture , In order to detect small and dense objects

(3)  Fuse the detection models of different sizes , namely Ensemble operation , Because the size of different targets may vary greatly , Such as seaports and ships 、 Airports and planes ,Ensemble Operation can improve the recognition accuracy under large size differences .

For satellites The image size is too large The problem of ,YOLT Cut into pieces , Cut the original image into small pieces and input them into the model for training , And combine 2-（3） Model fusion .

▲ YOLT Network structure , Most of the output feature sizes are 26 x 26, It can improve the detection accuracy

Application example

From the following detection example , We can see YOLT How it works ：

First , The development team adjusted a satellite image to 416 x 416 size （ Above left ）, It is found that the vehicle target cannot be detected ;

And cut from the original 416 x 416 Region （ Call it Chips） Then the detection of some vehicle targets can be realized

.

Follow this line , The development team cut the original image into many chips, And make adjacent chips There is a certain overlap between （ Pictured above ）, To ensure the integrity of image detection .

Again utilize NMS The algorithm will repeatedly detect and filter , Finally, the detection results of each block are fused , You can get the final result .

▲ Test example ： use YOLT v4 Identify the aircraft in the airport

YOLT The idea of can not only be applied to satellite image recognition , It can also play a role in other kinds of image recognition problems with small and dense target size .

Yes, this time Amazon 【AI For Good - 2022 Remote sensing optical image target detection challenge 】 For the contestants ,YOLT Our solution can help you overcome the first barrier of too large image size .

The competition registration will continue until 8 month 24 Japan , Students who have not participated in the competition still have enough time to prepare for the competition .

  Scan the QR code below or click to read the original   

  Sign up for   

▼

* Event benefits  *