Search

CN-122024156-A - Sea ice concentration recognition method based on DeepLab v3+ model vision technology

CN122024156ACN 122024156 ACN122024156 ACN 122024156ACN-122024156-A

Abstract

The invention discloses a sea ice density identification method based on DeepLab v & lt3+ & gt model vision technology, which relates to the technical field of ship navigation and comprises the following steps of S1, acquiring sea area video data around a ship in real time, processing the video data into an original frame, preprocessing the original frame to obtain a frame to be identified, and S2, inputting the frame to be identified into a DeepLab v & lt3+ & gt network to obtain sea ice density C. Compared with the prior art, the method has the advantages of high recognition precision and short training period.

Inventors

  • JIANG JINHUI
  • TANG WENYONG
  • YUAN YUCHAO
  • JIANG HERONG

Assignees

  • 上海交通大学
  • 上海船舶运输科学研究所有限公司

Dates

Publication Date
20260512
Application Date
20260126

Claims (10)

  1. 1. A sea ice density identification method based on DeepLab v3+ model vision technology is characterized by comprising the following steps: s1, acquiring the video data of the sea area around the ship in real time, processing the video data into an original frame, and preprocessing the original frame to obtain a frame to be identified; s2, inputting the frame to be identified into DeepLab v3+ network to obtain sea ice density C, The DeepLab v3+ network includes an encoder and a decoder, the encoder including a backbone network and a hole space pyramid pooling module.
  2. 2. The sea ice concentration identification method of claim 1, wherein the backbone network is Xception architecture, and the backbone network outputs three scale feature maps of 1/2, 1/4 and 1/8 to the decoder.
  3. 3. A sea ice concentration identification method as claimed in any one of claims 1 and 2 wherein the hole space pyramid pooling module includes a parallel convolution block attention module that is located before the hole convolution parallel layer of the pyramid pooling module and receives the output of the backbone network.
  4. 4. The sea ice intensity recognition method of claim 1, further comprising: step S3, result output and navigation system interface S301, outputting the C to a sea ice and ice load monitoring system; s302, when C is more than 60% and the ship speed is more than 8kn, sound and light early warning is sent out.
  5. 5. The sea ice intensity recognition method of claim 1, wherein the sea ice intensity C is calculated by: Removing sky area, counting the number of ice pixels N_ice and the number of water pixels N_water, Sea ice concentration c=n_ice/("a"); N_ice+N_water x 100%.
  6. 6. The sea ice concentration identification method of claim 1, wherein the DeepLab v3+ network training method comprises: Selecting 500 navigation images as a training data set; And (3) pre-training the model, namely selecting 108 sea ice pictures by adopting a method based on DeepLab semantic segmentation frame transfer learning, making a sea ice training set with real segmentation labels, loading initial weights of the Voc2012 dataset and retraining an output layer to obtain an identification model suitable for sea ice scenes, and selecting the precision of a sea ice test image verification model.
  7. 7. The method of claim 6, wherein, The navigation image is a side view ice image and covers 5 real scenes including crushed ice, gray ice, thick ice, open water surface and sky background; the training data set labeling specification is that an open source software tool is used for manually closing a contour, and a single-channel PNG label is output, wherein 0=seawater, 1=sea ice, 2=sky and 3=hull; the training data set is divided into 400 training pieces and 100 verification pieces, and random seeds are fixed.
  8. 8. A sea ice concentration recognition system comprising a camera and a central processing device arranged on a vessel, the sea ice concentration recognition system being adapted to implement the method of claim 1.
  9. 9. Sea ice concentration identification system as claimed in claim 8, wherein the number of cameras is 3, The first camera is arranged on the ship steering table, the direction is forward looking, The second camera is arranged on the deck of the driving platform, the direction is rearview, The third camera is arranged on the deck of the bow building and faces to the sea ice breaking place.
  10. 10. The sea ice concentration identification system of claim 8, wherein the camera is an arc-top type miniature camera, and the camera is connected with the central processing device by single-mode fiber wiring.

Description

Sea ice concentration recognition method based on DeepLab v3+ model vision technology Technical Field The invention relates to the technical field of ship navigation, in particular to a sea ice concentration identification method based on DeepLab v3+ model vision technology. Background The northeast channel of North China can shorten the European route by more than 40%, and the commercial ships in China in 2013-2024 safely complete more than 60 passes. With the acceleration of commercial operations of arctic airlines, the safety of vessels sailing in ice areas has become a critical issue. Sea ice concentration is one of the core parameters for evaluating navigation risk, and directly influences ship speed selection, route planning and structural safety. The traditional sea ice concentration acquisition means comprises: (1) Satellite remote sensing, namely low spatial resolution (250 m-1 km), long revisiting period and failure of cloud shielding; (2) The ship-borne radar is seriously affected by rain, snow and fog, and the echo has poor distinguishing degree of flat ice and boiled water; (3) The subjectivity is strong, the night/polar night can not work and can not be quantified by manual visual inspection. In the last five years, academia begins to try to realize ship-based visual ice condition monitoring by using a ship-based high-definition camera and an image segmentation algorithm. However, the global threshold value or local adaptive binarization method adopted generally can not meet the ice condition data requirement of IMO (in-field navigation rule) on 'real-time, quantifiable and error < 10%' under the conditions of low illumination, low contrast, snow fog and backlight of the polar region. Accordingly, those skilled in the art have been working to develop a sea ice concentration identification method. Disclosure of Invention In view of the above-mentioned drawbacks of the prior art, the technical problems to be solved by the present invention are: 1. the segmentation precision is low, and the traditional gray threshold method simplifies the image into two types of ice/water, and can not distinguish broken ice, grey ice, snow covered ice and reflective flare, so that the density is systematically overestimated or underestimated. 2. Edge loss-the ice-water boundary creates "jaggies" or "holes" when the backlight or crushed ice scale is less than 10 pixels, distorting the area statistics. 3. The real-time performance is poor, the existing algorithm (such as SVM-based pixel classification) needs manual design features, the calculation complexity is high, the single-frame processing time is more than 1s, and the real-time monitoring requirement of the ship 30fps can not be met. 4. Data loss, namely, the arctic sea ice image aiming at the view angle of a commercial ship is lacking in the public data set, and the model generalization capability is weak. In order to achieve the above purpose, the invention provides a sea ice concentration identification method based on DeepLab v3+ model vision technology, which comprises the following steps: s1, acquiring the video data of the sea area around the ship in real time, processing the video data into an original frame, and preprocessing the original frame to obtain a frame to be identified; s2, inputting the frame to be identified into DeepLab v3+ network to obtain sea ice density C, The DeepLab v3+ network includes an encoder and a decoder, the encoder including a backbone network and a hole space pyramid pooling module. Further, the backbone network is Xception architecture, and the backbone network outputs three scale feature graphs of 1/2, 1/4 and 1/8 to the decoder. Further, the hole space pyramid pooling module comprises a parallel convolution block attention module which is positioned in front of the hole convolution parallel layer of the pyramid pooling module and receives the output of the backbone network. Further, the method further comprises: Step S3, result output and navigation interface S301, outputting the C to a sea ice and ice load monitoring system; s302, when C is more than 60% and the ship speed is more than 8kn, sound and light early warning is sent out. Further, the method for calculating the sea ice concentration C is to exclude sky areas (sky types), count the number of pixels n_ice of ice (ice) and the number of pixels n_water of water (water), and calculate the sea ice concentration c=n_ice/(n_ice+n_water) ×100%. Further, the DeepLab v3+ network training method includes: Selecting 500 navigation images as a training data set; And (3) pre-training the model, namely selecting 108 sea ice pictures by adopting a method based on DeepLab semantic segmentation frame transfer learning, making a sea ice training set with real segmentation labels, loading initial weights of the Voc2012 dataset and retraining an output layer to obtain the identification model applicable to sea ice scenes. And selecting sea ice test images to verify the accuracy