Search

CN-122024115-A - Low-altitude remote sensing anomaly identification method based on artificial intelligence

CN122024115ACN 122024115 ACN122024115 ACN 122024115ACN-122024115-A

Abstract

The invention relates to a low-altitude remote sensing anomaly identification method based on artificial intelligence, belonging to the technical field of digital image processing and geographic information remote sensing monitoring. The method constructs a dynamic, content-aware feature learning and decision framework. Firstly, the parallel branches with different expansion rates are used for comprehensively capturing diversified receptive field features from local details to global contexts, then channel compression and excitation and pooling operation according to channel dimensions are respectively carried out on the diversified receptive field features, a lightweight channel-space two-way gating mechanism is realized, proper weights are adaptively distributed for feature graphs of different branches and different positions, geographical scene priori knowledge is fused, spatial attention is applied on a feature layer, background area response irrelevant to abnormality is further restrained, and model attention is focused on a real suspicious area. Finally, the detection capability of multi-scale anomalies, especially small-scale and weak-contrast anomalies, can be remarkably improved, so that the low-altitude remote sensing anomaly identification accuracy is improved.

Inventors

  • Tan Richang

Assignees

  • 北京吉威空间信息股份有限公司

Dates

Publication Date
20260512
Application Date
20260414

Claims (9)

  1. 1. An artificial intelligence-based low-altitude remote sensing anomaly identification method is characterized by comprising the following steps: Constructing an input feature map by any low-altitude remote sensing image, and calculating a preset variety of spectrum index maps of the input feature map; constructing a preset number of cavity convolution branches with the same convolution kernel size and different expansion rates, respectively inputting an input feature map into each cavity convolution branch to obtain a preset number of output feature maps, and splicing all the output feature maps according to channel dimensions to obtain multi-receptive-field features; channel compression and excitation are carried out on the multi-receptive field features to obtain channel attention weight vectors with the same number of dimensions as the number of the multi-receptive field channels, and the channel attention weight vectors and the multi-receptive field features are multiplied channel by channel to obtain channel weighting features; respectively carrying out maximum pooling and average pooling on the channel weighted features according to the channel dimension, splicing the maximum pooling result and the average pooling result, generating a space attention diagram through a standard convolution layer with a preset size, and multiplying the space attention diagram with the channel weighted features element by element to obtain a fusion weighted feature; Selecting a related spectrum index graph from preset types of spectrum index graphs of an input feature graph according to a current abnormality recognition task target, constructing a priori attention mask according to the related spectrum index graph, converting the priori attention mask into a priori attention weight graph, and enhancing fusion weighting features according to the priori attention weight graph to obtain an abnormality sensitive feature graph; and completing the abnormality identification under the current abnormality identification task target according to the abnormality sensitive characteristic diagram.
  2. 2. The artificial intelligence based low-altitude remote sensing anomaly identification method of claim 1, wherein the constructing an input feature map comprises: cutting any low-altitude remote sensing image to obtain a plurality of image blocks with first preset pixel numbers, wherein overlapping areas with second preset pixel numbers exist between adjacent image blocks, and any image block is used as the input feature map.
  3. 3. The artificial intelligence based low-altitude remote sensing anomaly identification method of claim 1, wherein the preset variety of spectral index maps comprise a normalized difference vegetation index map, a normalized difference water body index map, a normalized difference building index map and a bare earth surface index map.
  4. 4. The method for identifying low-altitude remote sensing anomalies based on artificial intelligence according to claim 1, wherein the performing channel compression and excitation on the multi-receptive field features to obtain channel attention weight vectors with the same number of dimensions as the number of the multi-receptive field channels comprises: Sequentially carrying out average pooling on each channel of the multi-receptive field features to obtain a corresponding average pooling result, and forming channel compression result vectors with the same number of dimensions as the multi-receptive field channels according to the average pooling results of all channels; Calculating the channel attention weight vector according to the channel compression result vector: , Wherein, the Representing the channel attention weight vector, Representing the channel compression result vector in question, Representing a first matrix of learnable parameters of the size of Wherein Representing the total number of convolved branches of the hole, The number of channels of the output characteristic diagram output by each cavity volume integral branch is represented, The reduction ratio is indicated as such, Representing the filtering function, acting to determine Whether each element in the calculation result is smaller than 0 or not, and when the element is smaller than 0, the corresponding element is replaced by 0, Representing a second matrix of learnable parameters of the size of , Representing the normalization function.
  5. 5. The artificial intelligence based low-altitude remote sensing anomaly identification method of claim 1, wherein the generating a spatial attention map by a standard convolution layer of a preset size after the maximum pooling result and the average pooling result are spliced comprises: And after the maximum pooling result and the average pooling result are spliced, inputting the spliced result into the standard convolution layer with the preset size, and normalizing the output of the standard convolution layer to obtain the space attention diagram.
  6. 6. The artificial intelligence based low-altitude remote sensing anomaly identification method of claim 1 or 5, wherein the convolution kernels of the standard convolution layers of the preset size are equal in length and width and are odd in number.
  7. 7. The artificial intelligence based low-altitude remote sensing anomaly identification method of claim 3, wherein constructing a prior attention mask from the correlation spectral index map comprises: The spectrum index graph which is inversely related to the current abnormality identification task target in the related spectrum index graph is marked as a negative correlation index graph, and the spectrum index graph which is positively related to the current abnormality identification task target in the related spectrum index graph is marked as a positive correlation index graph; the difference between the constant 1 and any negative correlation index map is recorded as an inverse mapping index map, and each inverse mapping index map and each positive correlation index map are multiplied element by element to obtain the prior attention mask.
  8. 8. The artificial intelligence based low-altitude remote sensing anomaly identification method of claim 1 or 7, wherein the converting into a priori attention weighting map comprises: And inputting the prior attention mask into a convolution layer with the length and the width of 1 to obtain a convolution output result, and normalizing the convolution output result to obtain the prior attention weight graph.
  9. 9. The artificial intelligence based low-altitude remote sensing anomaly identification method of claim 1, wherein the anomaly sensitivity feature map is: , Wherein, the Representing the abnormal sensitivity characteristic diagram, Representing a learnable a priori attention intensity control parameter, Representing the a priori attention weighting map, Representing an element-by-element multiplication, Representing the fusion weighting characteristics.

Description

Low-altitude remote sensing anomaly identification method based on artificial intelligence Technical Field The invention relates to the technical field of digital image processing and geographic information remote sensing monitoring, in particular to a low-altitude remote sensing anomaly identification method based on artificial intelligence. Background The low-altitude remote sensing technology mainly refers to a technology for obtaining remote sensing image data with high spatial resolution and high timeliness by utilizing platforms such as unmanned aerial vehicles, light airplanes and the like to observe the ground at a low altitude of hundreds to thousands of meters from the ground. The technology plays an irreplaceable role in the fields of homeland mapping, geographical information updating, mineral resource exploration and the like. By intelligently identifying the anomaly of the low-altitude remote sensing image, the change, anomaly or potential risk target of the earth surface can be efficiently found. Such as illegal mining sites, landslide hazards, water pollution diffusion, forest fire risk areas and the like. The technology converts the traditional visual interpretation method which depends on manpower and has long period and limited coverage into the efficient, objective and large-scale automatic monitoring method, and greatly improves the management efficiency and emergency response speed of related industries. Because the geographical scene covered by the low-altitude remote sensing image has great diversity from vast plains, undulating hills to steep mountain areas, the texture, spectrum and spatial structure of the ground surface covering (such as vegetation, water body, bare rock and artificial building) are complex and changeable, so that the target to be identified is required to be identified, and the scale variation range is extremely large. For example, in mineral resource development and supervision, it is necessary to identify large illegal mining surfaces with areas of possibly tens of square meters, and to locate new construction sheds or small illegal mining points with only tens of square meters. Therefore, the "multiscale" is a core feature of low-altitude remote sensing anomaly identification. Existing generic object detection models typically employ a fixed backbone network (e.g., resNet, VGG) in conjunction with a Feature Pyramid Network (FPN) to extract multi-scale features. However, such methods can significantly degrade their performance in the face of the extreme dimensional changes (especially small scale anomalies) and complex background disturbances described above. The first reason is that the feature extraction network of the fixed structure has limited characterization capability when facing small-scale abnormal features with large differences from the distribution of training data, especially with high similarity to the background or extremely low signal to noise ratio, and the downsampling operation of the deep network is extremely easy to cause the loss of key detail information of a small target. Second, complex natural scene backgrounds (e.g., dense vegetation shadows, bare rock textures, cloud cover, etc.) can produce a large number of interference features similar in color to real anomalies in shallow textures, and models can easily misinterpret these background noises as foreground anomalies, or otherwise, drown real anomalies embedded in complex backgrounds. The practical value and the credibility of the low-altitude remote sensing intelligent monitoring system in actual deployment are severely restricted by the multi-scale abnormality, especially the omission and the false detection of the small-scale low-contrast abnormality under the complex background. Therefore, a low-altitude remote sensing anomaly identification method capable of self-adapting to complex geographic scenes, accurately focusing on multi-scale anomaly characteristics and effectively inhibiting background interference is needed. Disclosure of Invention In view of the above, the invention provides a low-altitude remote sensing anomaly identification method based on artificial intelligence, which aims to solve the technical problems of high omission rate and weak background interference resistance of the current low-altitude remote sensing anomaly identification method on multiscale anomalies, particularly small-scale anomalies, in a complex geographic scene. The invention discloses a low-altitude remote sensing anomaly identification method based on artificial intelligence, which comprises the following steps: Constructing an input feature map by any low-altitude remote sensing image, and calculating a preset variety of spectrum index maps of the input feature map; constructing a preset number of cavity convolution branches with the same convolution kernel size and different expansion rates, respectively inputting an input feature map into each cavity convolution branch to obtain a pres