CN-121392578-B - Coastal infrared small target detection method based on frequency domain and airspace cooperation

CN121392578BCN 121392578 BCN121392578 BCN 121392578BCN-121392578-B

Abstract

The invention discloses a coast infrared small target detection method based on the cooperation of a frequency domain and a space domain, which comprises the following steps of 1) acquiring a coast infrared image data set and preprocessing to make any infrared image be The corresponding binary mask label is 2) Constructing a context guide including a sea clutter spectrum suppression module Infrared small target detection network of module and pair Processing to obtain predicted segmentation mask And 3) based on And And 4) optimizing network parameters by utilizing gradient descent to obtain a final infrared small target detection model for outputting a detection result. The method can realize accurate detection of the small target in the infrared image to be detected, thereby meeting the detection requirement of the infrared small target on the coast.

Inventors

SUN GUANGLING
ZHU CHONGBO
WANG ZIHAO
Hao Daokuo
ZHOU CHEN
HOU WANYING

Assignees

安徽建筑大学

Dates

Publication Date: 20260508
Application Date: 20251017

Claims (5)

1. A coast infrared small target detection method based on frequency domain and airspace cooperation is characterized by comprising the following steps of 1, acquiring a coast infrared image data set and preprocessing to obtain a preprocessed coast infrared image set and a binary mask label set corresponding to the preprocessed coast infrared image set, and marking any coast infrared image in the preprocessed coast infrared image as And (2) and The height of (2) is , Is of the width of Order-making machine Is marked as a binary mask label ; Step 2, constructing an infrared small target detection network, which comprises a sea clutter spectrum suppression module and a context guide Module and pair of Processing to obtain predicted segmentation mask ; Step 2.1, the sea clutter spectrum suppression module pair Processing to obtain spatial domain image with suppressed sea clutter ; Step 2.2, the contextual guidance The module comprises a context-guided encoder, a feature fusion decoder, and a context-guided decoder Processing to obtain predicted segmentation mask ; Step 2.2.1, context-guided encoder comprising CGBlock module, lightweight Encoder and pair of Processing to obtain a multi-scale hierarchical feature graph set ; Step 2.2.2, feature fusion decoder pair Processing to obtain a predicted segmentation mask Step 2.2.1.1, CGBlock modules include a local feature extractor Surrounding context extractor Joint feature extractor Global context extractor Step a, local feature extractor Using a pair of convolutions Processing to obtain local features ; Step b, surrounding context extractor Using the mth different expansion ratio Depth separable convolutional layer pair of (a) Processing to obtain the mth surrounding context feature ; Step c, in the channel dimension And (3) with After splicing, inputting the joint feature extractor And pass through a convolution layer and After the activation function is processed, the mth joint feature is obtained ; Step d, global context extractor Using a fully-connected layer After the number of channels of (a) is compressed, reuse is carried out Activating and processing the compressed features to obtain the mth compressed feature At the same time, using another full connection layer The channel number of the channel is restored to the original size to obtain the weight vector of the attention of the mth channel Thereby obtaining an mth channel enhancement feature using equation (1) ; (1) In the formula (1), the components are as follows, Is that The function of the function is that, Representing a channel-by-channel multiplication; Step 2.2.1.2 lightweight An encoder including a serialization and embedding layer, an L layer Coding unit and rearrangement unit, wherein each layer The coding unit comprises a multi-head self-attention mechanism layer and a feedforward neural network; step A, serialization and embedding layer will Dividing and flattening the image blocks into a plurality of image blocks, and mapping each image block into a feature vector through a linear projection layer to form an m-th image block embedding sequence And then will After adding the position codes to be learned, an mth code embedded sequence is formed ; Step B, Sequentially through layers L Stacking calculation of the coding units to obtain the mth coded characteristic sequence ; Step C, a rearrangement unit will The coding features corresponding to each image block in (a) are rearranged to correspond to The mth scale-level feature map E m , having the same width and height, yields a set of multi-scale-level feature maps { E m |m=1, 2,..m }, where, E M is the enhancement feature of the deepest level, and M represents the total scale level number; Step 3, based on And Construction of composite loss function Step 4, optimizing and updating all weight parameters of the infrared small target detection network by using a gradient descent algorithm, and obtaining a composite loss function Stopping training when the maximum training iteration times are converged or reached, and obtaining a final infrared small target detection model which is used for processing the infrared image to be detected and outputting a small target detection result.
2. The method for detecting a small infrared target on shore according to claim 1, wherein step 2.2.2 comprises: Step 2.2.2.1, initializing m=m, using E m as decoding feature D m of the M-th scale; Step 2.2.2, inputting the D m into a feature fusion decoder, and upsampling the D m through a transposed convolution layer to obtain an m-1 scale upsampled enhancement feature U m-1 ; Step 2.2.2.3, splicing the U m-1 and the m-1 scale level characteristic diagram E m-1 in the channel dimension through jump connection to obtain an m-1 scale spliced enhancement characteristic Cat m-1 ; Step 2.2.2.4, performing feature extraction and refining on Cat m-1 through a context guide block, and outputting m-1 scale decoding features D m-1 ; step 2.2.2.5, after assigning m-1 to m, returning to step 2.2.2.2 for sequential execution until m=1, thereby obtaining a final decoding feature map Thereafter, the data is passed through a convolution layer The number of channels is mapped to 1 and then passes through Processing of activation functions to generate Each pixel point of the two-dimensional image is a probability of a foreground object and forms a predicted binary segmentation mask 。
3. The method for detecting small infrared targets on shore according to claim 1, wherein in step 3, the composite loss function is constructed by using formula (2) : (2) In the formula (2), the amino acid sequence of the compound, And Is 2 super-parameter weight coefficients, and ; For the recovery loss in the frequency domain, For the space domain segmentation loss, the method comprises the following steps: (3) in the formula (3), the amino acid sequence of the compound, Is that The pixel intensity value of row i and column j, Is that Pixel intensity values for row i and column j; (4) In the formula (4), the amino acid sequence of the compound, As the weight coefficient of the light-emitting diode, Representing a binary cross entropy loss function, Representation of A loss function, incorporating: (5) In the formula (5), the amino acid sequence of the compound, Is that The pixel true category label of the ith row and jth column of the document, Is that The prediction probability of the ith row and jth column pixels; (6) In the formula (6), the amino acid sequence of the compound, Is a smoothing coefficient.
4. An electronic device comprising a memory and a processor, wherein the memory is configured to store a program that supports the processor to perform the coast infrared small object detection method of any one of claims 1-3, the processor being configured to execute the program stored in the memory.
5. A computer readable storage medium having stored thereon a computer program, characterized in that the computer program when executed by a processor performs the steps of the coast infrared small target detection method of any of claims 1-3.

Description

Coastal infrared small target detection method based on frequency domain and airspace cooperation Technical Field The invention belongs to the field of infrared small target detection, and particularly relates to a coastal infrared small target detection method based on frequency domain and airspace cooperation. Background In recent years, with the wide deployment of infrared imaging equipment in the scenes of coastal security, maritime supervision, unmanned aerial vehicle inspection and the like, infrared small target detection technology is rapidly developed. Compared with visible light imaging, the infrared image can work at night under the condition of low visibility and is not influenced by illumination change, so that the method has natural advantages for detecting small target tasks in a large visual field. However, the existing coast infrared image set has a series of problems of strong sea wave interference, small target size, low signal to noise ratio and the like, so that the target is difficult to detect. The traditional model driving method comprises a filtering method, a local contrast and saliency method, a low-rank and sparse representation method and the like, is simple in calculation and high in interpretation, but has high false alarm rate under the interference of various extreme weather and the like on the coast, and is sensitive to parameters. The data driving method based on deep learning comprises convolutional neural network detection and self-attention mechanismThe method has strong characteristic self-learning capability, but still faces the defects of scarcity of coastal infrared samples, high global attention calculation, limited detection precision caused by insufficient frequency domain information utilization and the like. Disclosure of Invention The invention aims to solve the defects of the prior art, and provides a coast infrared small target detection method based on the cooperation of a frequency domain and a space domain, so that the frequency domain information of an image can be fully utilized, the signal to noise ratio of the image is reduced, the efficient multi-scale context fusion and the lightweight detection are realized, the accuracy of small target detection in an infrared image can be improved, and the coast infrared small target detection requirement can be met. In order to achieve the aim of the invention, the invention adopts the following technical scheme: the invention discloses a coast infrared small target detection method based on frequency domain and airspace cooperation, which is characterized by comprising the following steps of: step 1, acquiring a coast infrared image data set and preprocessing to obtain a preprocessed coast infrared image set and a binary mask label set corresponding to the preprocessed coast infrared image set, wherein any coast infrared image in the preprocessed coast infrared image is recorded as And (2) andThe height of (2) is,Is of the width ofOrder-making machineIs marked as a binary mask label; Step 2, constructing an infrared small target detection network, which comprises a sea clutter spectrum suppression module and a context guideModule and pair ofProcessing to obtain predicted segmentation mask; Step 2.1, the sea clutter spectrum suppression module pairProcessing to obtain spatial domain image with suppressed sea clutter; Step 2.2, the contextual guidanceThe module comprises a context-guided encoder, a feature fusion decoder, and a context-guided decoderProcessing to obtain predicted segmentation mask; Step 2.2.1, context-guided encoder comprising CGBlock module, lightweightEncoder and pair ofProcessing to obtain a multi-scale hierarchical feature graph set; Step 2.2.2, feature fusion decoder pairProcessing to obtain a predicted segmentation mask; Step 3, based onAndConstruction of composite loss function; Step 4, optimizing and updating all weight parameters of the infrared small target detection network by using a gradient descent algorithm, and obtaining a composite loss functionStopping training when the maximum training iteration times are converged or reached, and obtaining a final infrared small target detection model which is used for processing the infrared image to be detected and outputting a small target detection result. The coast infrared small target detection method of the invention is also characterized in that the step 2.1 comprises the following steps: step 2.1.1, pair Performing two-dimensional fast Fourier transform to obtain complex frequency domain representation; Step 2.1.2, calculatingAmplitude spectrum of (a)And phase spectrum; Step 2.1.3, willFront in the middle-height directionFront of each pixel in width directionThe region formed by the pixels is denoted as low frequency band; Will beThe first in the middle-height directionThe first pixel to the second pixelFirst pixel in width directionThe first pixel to the second pixelThe region formed by each pixel is denoted as high-frequency bandWh