An automatic classification method for space-based observation images

Abstract

The application relates to the technical field of digital image processing and pattern recognition, and discloses an automatic classification method for space-based observation images, which comprises the following steps: acquiring original space-based observation images by a space-based observation load and converting the original space-based observation images into an original feature matrix, and when the local contrast of the original feature matrix is lower than 20% of a preset global reference variance, the first regulation and control smoothing constant of a sliding window region is adjusted upward; a second-order difference operator is used to extract a feature matrix variation spectrum density, an anisotropic operator is called to obtain a bidirectional gradient histogram component to construct an asymmetric scaling tensor; an intrinsic scale scaling coefficient is derived, and the shear deformation rate of an affine transformation matrix is corrected to construct a transformation matrix; the original feature matrix is resampled to output a feature normalization matrix, and spatial pooling and mapping are performed, and a classification result is output; and the application maintains the geometric topological stability of the front scale alignment through the non-gated continuous feature alignment.

Description

Technical Field

[0001] This invention relates to an automatic classification method for space-based observation images, belonging to the field of digital image processing and pattern recognition technology. Background Technology

[0002] Currently, image classification typically utilizes deep neural networks to extract spatial structural features from digital image matrices and outputs classification decisions through target category probability mapping models. However, for space-based observation platforms, the relative distance between the optical imaging payload and the ground target, as well as the field of view, continuously change during operation. This leads to significant scale variations and spatial geometric distortions in the acquired two-dimensional digital image raster. Furthermore, non-uniform atmospheric transport and sudden dense fog can cause rapid changes in high-frequency edge information in the image, resulting in decreased edge contrast and texture feature confusion. These degradations affect the stability of data distribution in the initial feature extraction stage, consequently shifting the probability vector output by the subsequent classification decision layer and impacting the reliability of the image classification results. To reduce spatial distortion... The impact of scale changes and weather degradation on classification accuracy is addressed by existing solutions, which typically involve setting multiple parallel convolutional kernels with different receptive fields within the network or constructing multi-level downsampled image pyramids before feature extraction. These methods essentially adapt to scale changes and geometric deformations in images by increasing spatial computation, which can significantly increase feature dimensions, leading to increased data processing latency and memory usage. When the computing power and power consumption of edge computing nodes are limited, the system is prone to problems such as excessive computing load and slow convergence of classification probability vectors. Furthermore, the discrete hard threshold gating switching channels or global mean equalization processing methods used in existing solutions can easily cause high-frequency jitter in the classifier's decision state near critical data boundaries and are also difficult to adapt to the local feature loss caused by non-uniform fog.

[0003] Besides being constrained by the hardware limitations of the underlying computing power and power consumption of edge processing nodes, existing image correction and preprocessing algorithms also have significant fundamental shortcomings when dealing with large-angle non-orthophoto deformations. For example, Chinese invention patent CN117152220B discloses a method for registering large-angle aerial images with orthophoto images. It uses the ASIFT algorithm to simulate multi-view affine images and extracts feature points in nonlinear scale space using phase consistency and the Sobel operator to complete image correction. However, this method is highly dependent on dense view sampling and the generation of massive simulated sub-images, introducing extremely high computational complexity and data redundancy in scale space construction and feature descriptor solving. This is incompatible with the low latency and lightweight constraints required by airborne edge nodes. In addition, its feature extraction mechanism is specifically optimized for geometric registration between multi-source images. If it is directly deployed in the classification front end, it will not only fail to spontaneously tune the classification feature weights, but will also cause the classifier decision vector to converge slowly due to complex matrix inverse sampling reconstruction.

[0004] Therefore, the technical problem to be solved by this invention is how to provide an automatic classification method for space-based observation images that can spontaneously tune geometric transformation weights and eliminate multi-scale architecture redundancy during the pre-feature extraction stage. Summary of the Invention

[0005] To address the problems in the background art, the technical solution of the present invention is as follows: An automatic classification method for space-based observation images, comprising the following steps: Step S1: Obtain the original space-based observation image from the space-based observation payload and convert it into the original feature matrix. Calculate the local contrast and gray-level variance of the original feature matrix. When the local contrast is lower than 20% of the preset global reference variance, increase the first control smoothing constant corresponding to the sliding window region proportionally. Step S2: Combine the first adjustment smoothing constant after the adjustment, use the second-order difference operator to extract the characteristic matrix change spectral density of the original characteristic matrix, and at the same time call the anisotropic operator to obtain the bidirectional gradient histogram components, and construct an asymmetric scaling tensor based on the ratio of the bidirectional gradient histogram components. Step S3: Use the spectral density of the characteristic matrix to derive the eigenscale scaling factor for global smooth convergence, and correct the shear deformation rate of the two-dimensional affine transformation matrix in different coordinate axis directions based on the asymmetric scaling tensor. Step S4: Construct a two-dimensional affine transformation matrix using the intrinsic scale scaling factor and the corrected shear deformation rate, resample the original feature matrix, output a feature normalization matrix with consistent specifications, perform pooling and vector mapping on the feature normalization matrix in a single-scale space, and output the final decision confidence and classification result.

[0006] Preferably, in step S2, the extraction of the feature matrix change spectral density of the original feature matrix using the second-order difference operator includes the following sub-steps: Step S21, calculate the second-order derivative matrix of the original feature matrix in the horizontal and vertical directions using the second-order difference operator; Step S22, perform Gaussian smoothing filtering on the second-order derivative matrix in combination with the up-adjusted first control smoothing constant to obtain the feature matrix change spectral density.

[0007] Preferably, in step S2, the process of calling the anisotropic operator to obtain the bidirectional gradient histogram components includes the following sub-steps: Step S23, using the anisotropic operator to calculate the gradient intensity and gradient direction along the horizontal and vertical axes of the original feature matrix; Step S24, statistically analyzing the gradient intensity distribution along different axes, constructing the horizontal gradient histogram components and the vertical gradient histogram components respectively, and combining them to form the bidirectional gradient histogram components.

[0008] Preferably, in step S3, correcting the shear deformation rate of the two-dimensional spatial affine transformation matrix in different coordinate axis directions based on the asymmetric scaling tensor includes the following sub-steps: Step S31, calculate the numerical ratio of the horizontal gradient histogram component to the vertical gradient histogram component to determine the perspective distortion under the non-orthographic projection form; Step S32, fill the off-diagonal elements of the asymmetric scaling tensor according to the perspective distortion, and use the off-diagonal elements to correct the shear deformation rate of the two-dimensional spatial affine transformation matrix in different coordinate axis directions to eliminate nonlinear perspective stretching deformation.

[0009] Preferably, in step S3, determining the intrinsic scaling factor using the spectral density of the eigenma matrix variation includes the following sub-steps: Step S33, calculating the range between the gradient maxima and gradient minima of the spectral density of the eigenma matrix variation; Step S34, when the high-frequency eigenenergy drops and the range approaches the zero singularity, triggering a monotonically complementary damped convergence constraint mechanism to stabilize the intrinsic scaling factor within the preset smooth convergence stable region, thus blocking the propagation of numerical divergence to the two-dimensional affine transformation matrix.

[0010] Preferably, in step S4, resampling the original feature matrix and outputting a feature normalization matrix with consistent specifications includes the following sub-steps: Step S41, using the two-dimensional affine transformation matrix constructed by the intrinsic scale scaling coefficient and the corrected shear deformation rate, a bidirectional geometric mapping relationship is established from the normalized single-scale space to the original image coordinate space; Step S42, based on the bidirectional geometric mapping relationship, the original feature matrix is ​​resampled to eliminate the geometric deformation caused by the change in observation height and the change in shooting field of view in the front end of the classification network, and a feature normalization matrix is ​​generated.

[0011] Preferably, in step S4, pooling and vector mapping of the feature normalization matrix in a single-scale space to output the final decision confidence and classification result includes the following sub-steps: Step S43, inputting the feature normalization matrix into a single-scale lightweight classification network, pooling and reducing the dimensionality of each channel of the feature normalization matrix in a single-scale space to construct intrinsic classification feature vectors; Step S44, using fully connected layers and classification mapping operators to map the intrinsic classification feature vectors to target class probabilities, and outputting the decision confidence and classification result.

[0012] Preferably, in step S43, pooling is performed on each channel of the feature normalization matrix in a single-scale space to reduce dimensionality and construct the intrinsic classification feature vector. This includes: jointly using global average pooling and global max pooling to compress the spatial dimension of the feature normalization matrix, extracting the average energy distribution and peak abrupt texture features of each feature channel, and splicing them together to generate the intrinsic classification feature vector.

[0013] Preferably, before calculating the local contrast and grayscale variance of the original feature matrix, a pre-observation parameter correction step is also included: Step S5, acquiring real-time remote sensing altitude data and barometric altitude data of the airborne observation payload, calculating the complementary deviation between the two, and using the complementary deviation to correct the initial spatial resolution benchmark of the original feature matrix, providing a geometric benchmark constraint for determining the intrinsic scale scaling factor.

[0014] Compared with the prior art, the beneficial effects of the present invention are: 1. In the automatic classification of space-based observation images, the maximum, minimum and average gradients in the spectral density of the feature matrix reflect the changes in image edges and textures. When the edges are blurred due to factors such as fog and backlight, the intrinsic scale scaling factor can be adjusted continuously. This can avoid the classification results jitter caused by hard threshold switching in the critical state, limit the abnormal divergence of scale coefficients, reduce the risk of transformation matrix instability and image tearing, and make the pre-scale alignment process more stable.

[0015] 2. By constructing a two-dimensional affine transformation matrix through intrinsic scale scaling coefficients and resampling the original feature matrix, scale normalization can be completed before the image enters the classification network, outputting a feature normalization matrix with consistent specifications. This can preprocess the geometric deformation caused by changes in observation height, enabling the subsequent classification process to complete pooling and vector mapping in a unified single-scale space. This reduces feature redundancy, computational consumption, and memory occupation caused by multi-scale image pyramids and multi-branch convolutions, making it more suitable for deployment on edge computing nodes with limited computing power.

[0016] 3. By constructing an asymmetric scaling tensor through bidirectional gradient histogram components and correcting the shear deformation rate of the two-dimensional affine transformation matrix in different coordinate axis directions, it is possible to correct the stretching and deflection of images under tilted viewpoints. Compared with the method of only performing scale compensation in a single direction, this processing can better adapt to multi-axial deformation in space-based observation images, so that the feature-aligned image maintains a more stable geometric relationship and reduces the interference of perspective deformation on classification decisions. Attached Figure Description

[0017] Figure 1 This is a flowchart of the feature extraction and affine transformation normalization process of the present invention; Figure 2 This is a diagram showing the switching of multiple operating states of the system of the present invention.

[0018] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0020] An automatic classification method for space-based observation images includes the following steps: Step S1: Obtain the original space-based observation image from the space-based observation payload and convert it into the original feature matrix. Calculate the local contrast and gray-level variance of the original feature matrix. When the local contrast is lower than 20% of the preset global reference variance, increase the first control smoothing constant corresponding to the sliding window region proportionally. Step S2: Combine the first adjustment smoothing constant after the adjustment, use the second-order difference operator to extract the characteristic matrix change spectral density of the original characteristic matrix, and at the same time call the anisotropic operator to obtain the bidirectional gradient histogram components, and construct an asymmetric scaling tensor based on the ratio of the bidirectional gradient histogram components. Step S3: Use the spectral density of the characteristic matrix to derive the eigenscale scaling factor for global smooth convergence, and correct the shear deformation rate of the two-dimensional affine transformation matrix in different coordinate axis directions based on the asymmetric scaling tensor. Step S4: Construct a two-dimensional affine transformation matrix using the intrinsic scale scaling factor and the corrected shear deformation rate, resample the original feature matrix, output a feature normalization matrix with consistent specifications, perform pooling and vector mapping on the feature normalization matrix in a single-scale space, and output the final decision confidence and classification result.

[0021] Preferably, in step S2, the extraction of the feature matrix change spectral density of the original feature matrix using the second-order difference operator includes the following sub-steps: Step S21, calculate the second-order derivative matrix of the original feature matrix in the horizontal and vertical directions using the second-order difference operator; Step S22, perform Gaussian smoothing filtering on the second-order derivative matrix in combination with the up-adjusted first control smoothing constant to obtain the feature matrix change spectral density.

[0022] Preferably, in step S2, the process of calling the anisotropic operator to obtain the bidirectional gradient histogram components includes the following sub-steps: Step S23, using the anisotropic operator to calculate the gradient intensity and gradient direction along the horizontal and vertical axes of the original feature matrix; Step S24, statistically analyzing the gradient intensity distribution along different axes, constructing the horizontal gradient histogram components and the vertical gradient histogram components respectively, and combining them to form the bidirectional gradient histogram components.

[0023] Preferably, in step S3, correcting the shear deformation rate of the two-dimensional spatial affine transformation matrix in different coordinate axis directions based on the asymmetric scaling tensor includes the following sub-steps: Step S31, calculate the numerical ratio of the horizontal gradient histogram component to the vertical gradient histogram component to determine the perspective distortion under the non-orthographic projection form; Step S32, fill the off-diagonal elements of the asymmetric scaling tensor according to the perspective distortion, and use the off-diagonal elements to correct the shear deformation rate of the two-dimensional spatial affine transformation matrix in different coordinate axis directions to eliminate nonlinear perspective stretching deformation.

[0024] Preferably, in step S3, determining the intrinsic scaling factor using the spectral density of the eigenma matrix variation includes the following sub-steps: Step S33, calculating the range between the gradient maxima and gradient minima of the spectral density of the eigenma matrix variation; Step S34, when the high-frequency eigenenergy drops and the range approaches the zero singularity, triggering a monotonically complementary damped convergence constraint mechanism to stabilize the intrinsic scaling factor within the preset smooth convergence stable region, thus blocking the propagation of numerical divergence to the two-dimensional affine transformation matrix.

[0025] Preferably, in step S4, resampling the original feature matrix and outputting a feature normalization matrix with consistent specifications includes the following sub-steps: Step S41, using the two-dimensional affine transformation matrix constructed by the intrinsic scale scaling coefficient and the corrected shear deformation rate, a bidirectional geometric mapping relationship is established from the normalized single-scale space to the original image coordinate space; Step S42, based on the bidirectional geometric mapping relationship, the original feature matrix is ​​resampled to eliminate the geometric deformation caused by the change in observation height and the change in shooting field of view in the front end of the classification network, and a feature normalization matrix is ​​generated.

[0026] Preferably, in step S4, pooling and vector mapping of the feature normalization matrix in a single-scale space to output the final decision confidence and classification result includes the following sub-steps: Step S43, inputting the feature normalization matrix into a single-scale lightweight classification network, pooling and reducing the dimensionality of each channel of the feature normalization matrix in a single-scale space to construct intrinsic classification feature vectors; Step S44, using fully connected layers and classification mapping operators to map the intrinsic classification feature vectors to target class probabilities, and outputting the decision confidence and classification result.

[0027] Preferably, in step S43, pooling is performed on each channel of the feature normalization matrix in a single-scale space to reduce dimensionality and construct the intrinsic classification feature vector. This includes: jointly using global average pooling and global max pooling to compress the spatial dimension of the feature normalization matrix, extracting the average energy distribution and peak abrupt texture features of each feature channel, and splicing them together to generate the intrinsic classification feature vector.

[0028] Preferably, before calculating the local contrast and grayscale variance of the original feature matrix, a pre-observation parameter correction step is also included: Step S5, acquiring real-time remote sensing altitude data and barometric altitude data of the airborne observation payload, calculating the complementary deviation between the two, and using the complementary deviation to correct the initial spatial resolution benchmark of the original feature matrix, providing a geometric benchmark constraint for determining the intrinsic scale scaling factor.

[0029] Example 1: The method in this example is deployed in an UAV airborne remote sensing imaging system. During the UAV's cruise operation, it acquires ground grid images through an airborne observation payload. The flight altitude and non-orthogonal shooting field of view change with the flight path, and the relative geometric distance between the imaging payload and the ground target fluctuates, causing nonlinear scaling of the same target pixel area in the digital image matrix. At the same time, local dense fog and strong backlighting will cause degradation of the underlying imaging, blurring the high-frequency edges of the image and causing a decrease in gradient energy. After receiving the original airborne observation image, the image processing unit converts it into the original feature matrix.

[0030] After conversion to the original feature matrix and before statistical analysis of grayscale variance, the pre-observation parameter correction step reads the real-time remote sensing altitude data and barometric altitude data uploaded by the UAV data bus through a digital signal processor. The two data are subtracted to obtain the absolute difference, which is used as the complementary bias. The digital signal processor multiplies the complementary bias by a preset geometrical projection coefficient to obtain the correction increment of the initial spatial resolution. This correction increment is then superimposed on the original initial spatial resolution reference to complete the online update of the resolution reference. The updated geometric reference constrains the physical size mapping during subsequent sliding window extraction of each feature layer, providing a unified geometric reference constraint for the determination of the intrinsic scale scaling factor.

[0031] The digital signal processor calls a sliding window to statistically analyze the gray-level variance and local contrast of each channel in the original feature matrix. When the local contrast within the sliding window is lower than 20% of the preset global baseline variance, the calculation unit adjusts the first control smoothing constant corresponding to that sliding window region upwards according to a preset proportional adjustment relationship. This improves the smoothing constraint on low-contrast regions in subsequent filtering processes. The proportional adjustment relationship uses the ratio of local contrast to global reference variance as input, with an initial smoothing constant reference of 1 as the reference value, and determines the upward adjustment amount according to the decrease of this ratio relative to the 20% boundary. When the local contrast drops to 15% of the global reference variance, the hardware interrupt controller activates the corresponding divider calculation path to calculate the first control smoothing constant of the output. The value is 1.25, which allows the smooth cutoff frequency to be adjusted according to the degree of degradation of the preceding image.

[0032] The corrected first regulation smoothing constant Under constraints, the spatial second-order difference operator calculates the second-order derivative matrix along the horizontal and vertical orthogonal directions of the original characteristic matrix. The computational unit then adjusts the first smoothing constant. As a smoothing parameter for Gaussian smoothing filtering, the second derivative matrix is ​​filtered to obtain the characteristic matrix change spectral density characterizing the high-frequency energy distribution in the spatial domain. At the same time, the anisotropic operator extracts the gradient intensity and gradient direction along the horizontal and vertical axes of the original characteristic matrix, and statistically analyzes the gradient intensity distribution along different axes. The horizontal gradient histogram component and the vertical gradient histogram component are constructed respectively. The computing unit divides the numerical mean of the horizontal gradient histogram component and the numerical mean of the vertical gradient histogram component to obtain the numerical ratio between the two. When the ratio deviates from 1, the absolute value of the deviation is determined as the perspective deflection under the non-orthophoto projection form.

[0033] When constructing the asymmetric scaling tensor, the computational unit keeps the diagonal elements of the asymmetric scaling tensor fixed at the initial reference value of 1, and fills the calculated perspective deflection into the off-diagonal elements in the lower left and upper right corners of the matrix to complete the quantization construction of the tensor matrix. These off-diagonal elements are used to correct the shear deformation rate of the affine transformation matrix in two-dimensional space in different coordinate axis directions, thereby eliminating the nonlinear perspective stretching deformation caused by the tilted shooting field of view.

[0034] The feature estimation unit performs local sliding window statistics on the spectral density of the feature matrix changes, dynamically obtaining the maximum gradient value in each sub-region. Minimum gradient value and average gradient value Furthermore, across the entire runtime domain, continuous computation logic replaces hard threshold logic branch switching, directly based on the maximum gradient value. Minimum gradient value With average gradient value Calculating the relative structural complexity factor using algebraic relationships in spatial gradient distribution. Its calculation formula is expressed as follows ,in, As a relative structural complexity factor, The average gradient value, The maximum gradient value, The minimum gradient value, This is a reference noise smoothing constant fixed in the storage unit, used to align with the physical limit of the underlying thermal noise to prevent the denominator from reaching absolute zero when the image is extremely smoothed. Image attenuation caused by atmospheric transmission degradation and strong backlighting manifests as a continuous transfer of characteristic high-frequency energy to mid- and low-frequency frequencies; at this time, the gradient range in the feature matrix is ​​greater than the average gradient value. Faster contraction drives the relative structural complexity factor The value of increases continuously towards 1. Under extremely blurred conditions, as high-frequency textures disappear, local areas of the image tend to become uniform gray grids, and the range between the maximum and minimum gradient values ​​rapidly narrows and approaches zero. The average gradient value within the region... It also decays synchronously; when the range shrinks to near the zero singularity, the average gradient value... It is close to the reference noise smoothing constant in numerical order. This ensures a computable dynamic correspondence between the numerator and denominator, constraining the relative structural complexity factor. It converges stably to around 1, preventing numerical divergence caused by denominator contraction.

[0035] Feature estimation units utilize relative structural complexity factors As a continuous tuning parameter, the intrinsic scale scaling factor is used to combat high abrupt changes. The formula for calculating ungated continuous regression is as follows: (This is a partial translation of the original text, which is not possible without further context.) ,in, This is the intrinsic scaling factor. As a relative structural complexity factor, The first preset smoothing constant is... The maximum gradient value, The average gradient value, It is a microconstant used to prevent the denominator from collapsing; For weight parameters, The minimum gradient value, The second anti-interference damping coefficient, which is fixed, exhibits a significant gradient spatial range and a relative structural complexity factor when the airborne remote sensing environment is under standard operating conditions and has rich surface texture details. Automatic convergence approaches 0, causing the solution weights of the unified regression formula to favor the geometric scaling factor solution term, thus aligning the geometric deformation across space. In the event of weather degradation or low-contrast conditions, the gradient's high-frequency energy decays, and the range closure reduces the relative structural complexity factor. As the value steadily approaches 1, the solution weights shift to the later-stage damping hedging term. Limit the intrinsic scale scaling factor The convergence extrema are obtained to reduce the risk of transformation matrix divergence and image tearing caused by high-frequency feature energy collapse.

[0036] The spatial transformation unit obtains the intrinsic scale scaling factor for global smooth convergence. Combined with the shear deformation rate corrected by the asymmetric scaling tensor, a two-dimensional spatial affine transformation matrix is ​​dynamically reconstructed. This two-dimensional spatial affine transformation matrix simultaneously records the coordinate mapping from the normalized single-scale space to the original image coordinate space, as well as the backsampling relationship from the original image coordinate space back to the normalized single-scale space. During resampling, the reconstruction transformation operator initiates inverse spatial bilinear interpolation resampling of the original feature matrix based on this mapping relationship. In the matrix processing process, the geometric deformation caused by the change in observation height and the change in shooting field of view is eliminated. In the pre-stage of image feature extraction, a feature normalization matrix with consistent scale space specifications is output.

[0037] Subsequently, the graphics computing chip directly inputs the feature normalization matrix into the single-scale lightweight classification network. The cascaded feature mapping operator jointly uses global average pooling and global max pooling to compress the spatial dimension of the feature normalization matrix, extracts the average energy distribution and peak abrupt texture features of each feature channel, and concatenates them to generate intrinsic classification feature vectors. The fully connected layer at the output end and the classification mapping operator then map the intrinsic classification feature vectors to target class probabilities, outputting the decision confidence and the surface target class label. This enables the classifier to maintain a stable response when continuously and gradually experiencing weather degradation and flying over desert areas lacking texture features, and reduces the computational redundancy and decrease in classifier confidence caused by the multi-scale architecture.

[0038] Example 2: This example verifies the automatic classification method for airborne observation images on an airborne remote sensing image verification platform. The verification platform simulates the remote sensing imaging environment of a UAV dynamically changing its cruising altitude and field of view through a digital signal generation unit. The test data originates from a sequence of ground target raster images collected by a physical experimental platform. Gaussian white noise with a signal-to-noise ratio of 20dB and power frequency interference harmonics with a frequency of 50Hz are superimposed in the signal source to create test conditions of low contrast and meteorological degradation superimposed. The first adjustment smoothing constant... Based on the characteristics of local grayscale changes, when the local contrast within the sliding window falls below 20% of the preset global baseline variance, the calculation unit adjusts the relationship according to a preset ratio to increase the first control smoothing constant. The value of ; under typical meteorological degradation conditions, when the local contrast drops to 15% of the global baseline variance, the first control smoothing constant is... The engineering setting value is 1.25.

[0039] Under the standard condition where the ratio of local contrast to global baseline variance is 85%, the gradient range of the control group is 112, and the average gradient is 24. After conventional processing, the relative structural complexity factor is obtained. The intrinsic scale scaling factor is 0.17. The relative structural complexity factor is 0.82, and the classification accuracy confidence level is 0.91. Under the same input conditions, the gradient range of the sample group of this invention is 114, and the average gradient value is 25. After processing according to the method of this embodiment, the relative structural complexity factor is... The intrinsic scale scaling factor is 0.18. With a score of 0.84 and a classification accuracy confidence score of 0.92, this data indicates that when the image texture is relatively intact, the intrinsic scale scaling factor... The gradient distribution remains close to the normal scale alignment, and the confidence level of classification accuracy does not show a significant decrease.

[0040] When the ratio of local contrast to global baseline variance drops to 50%, the gradient range of the control group is 64, the average gradient is 18, and the relative structural complexity factor is... The intrinsic scale scaling factor is 0.22. The confidence level is 0.65, and the classification accuracy confidence level is 0.76. The gradient range of the sample group under the same working conditions is 66, and the average gradient value is 19. After feature balancing, the relative structural complexity factor is... The intrinsic scale scaling factor is 0.23. The confidence level was 0.71, and the classification accuracy was 0.89. As local contrast decreased, the sample group of this invention passed the relative structural complexity factor test. Adjusting the intrinsic scale scaling factor This ensures that subsequent classification calculations are still performed on a relatively stable feature normalization matrix.

[0041] When the ratio of local contrast to global baseline variance further decreased to 20%, the gradient range of the control group shrank to 32, the average gradient value was 14, and the relative structural complexity factor... The intrinsic scale scaling factor is 0.31. The confidence level dropped to 0.42, with a classification accuracy of 0.51. The gradient range of the sample group under the same boundary conditions was 34, the average gradient was 15, and the relative structural complexity factor was... The intrinsic scale scaling factor is 0.32. The value is 0.64, and the classification accuracy confidence level is 0.88. This operating condition corresponds to the first control smoothing constant. When the working boundary is adjusted upwards, the gradient range in the spectral density of the eigenma matrix can still be used as the intrinsic scale scaling factor. Provide effective constraints.

[0042] Under the degradation condition where the ratio of local contrast to global baseline variance is 10%, the gradient range of the control group is 15, the average gradient is 11, and the relative structural complexity factor is... The intrinsic scale scaling factor is 0.44. The confidence level dropped to 0.18, and the classification accuracy confidence level dropped to 0.32; the gradient range of the sample group in this invention was 17, the average gradient value was 12, and the relative structural complexity factor was... The intrinsic scale scaling factor is 0.46. The convergence rate was 0.58, and the classification accuracy confidence level remained at 0.85. Compared with the gradient range of 114 and the average gradient of 25 for the sample group under standard conditions, when the gradient range shrank from 34 to 17 under this condition, the range collapse rate was faster than the decay rate of the average gradient. This is relative to the structural complexity factor. The intrinsic scale scaling factor then increases from 0.18 to 0.46. The solution centroid shifts to a damping hedging term, suppressing numerical divergence and image tearing of the two-dimensional affine transformation matrix.

[0043] When the ratio of local contrast to global baseline variance drops to 5%, the gradient range of the sample group in this invention is 8, the average gradient is 9, and the relative structural complexity factor is... The intrinsic scale scaling factor is 0.56. The ratio is 0.52, and the classification accuracy confidence level is 0.82. When this ratio further decreases to 2%, the gradient range of the sample group of this invention is 4, the average gradient value is 7, and the relative structural complexity factor is... The intrinsic scale scaling factor is 0.68. The intrinsic scale scaling factor of the control group under the recorded operating conditions was 0.49, with a classification accuracy confidence level of 0.79. As the confidence level continuously dropped from 85% to 2%, the intrinsic scale scaling factor of the control group under the recorded operating conditions was... The classification accuracy confidence level decreased from 0.91 to 0.32 as the variance decreased from 0.82 to 0.18. Under saturation conditions of 5% and 2%, the classification accuracy confidence level of the sample group remained at 0.82 and 0.79 respectively, indicating a gradual slowdown in the rate of decline. This suggests that when the local contrast is below 20% of the preset global baseline variance, the first adjustment smoothing constant... The upward adjustment can bring the feature balancing mechanism into the effective working range.

[0044] Experimental results show that, under non-ideal working conditions where Gaussian white noise and power frequency interference coexist, the method in this embodiment transforms spatial scale stretching deformation into a continuous evolution constraint of feature gradient through geometric alignment and feature balancing in the pre-feature extraction stage. The normalized feature normalization matrix can still maintain the classification feature expression without relying on multi-scale parallel convolution branches, thus alleviating the conflict between the low computing power constraint of edge processing nodes and the dimensional expansion of multi-scale feature matrices.

[0045] Example 3: In this example, the embedded digital signal processor in the image processing unit receives the original airborne observation image collected by the UAV optical payload in the strong backlight temperature-controlled cabin, and converts the original airborne observation image into an original feature matrix containing a pixel matrix through the internal first asynchronous transceiver bus. The first static random access memory inside the embedded digital signal processor allocates a sliding window data block for temporarily storing the current processing frame. The spatial size of the sliding window data block is 3 pixels by 3 pixels, and it slides and samples row by row along the horizontal and vertical directions of the original feature matrix with a step size of 1 pixel.

[0046] Within each sliding step, the arithmetic logic unit reads the grayscale values ​​of 9 pixels within the sliding window data block, sums them, and divides by 9 to obtain the average grayscale value of the window. Then, it subtracts the average grayscale value of the window from the grayscale value of each pixel to generate a zero-mean residual sequence. The arithmetic logic unit calculates the sum of squares of each element in the zero-mean residual sequence and divides the sum of squares by 8 to obtain the local contrast of the current area. At the same time, it compares the local contrast with the global reference variance stored in the second memory. When the local contrast drops to 15% of the global reference variance, the hardware interrupt controller triggers a state flip signal, driving the bus controller to increase the value written to the cutoff frequency register of the spatial smoothing window from the initial 1 to 1.25, so that the smoothing cutoff frequency is adjusted according to the degree of degradation of the previous imaging.

[0047] The feature estimation unit resamples the spectral density of the feature matrix after processing by the second-order spatial difference operator. The data is then transferred to the dedicated adder and subtractor array of the arithmetic logic unit. The adder and subtractor array performs parallel extremum scanning on the spatial gradient sequence within the current feature matrix to lock the current local maximum gradient value. with minimum gradient value Simultaneously, all spatial gradient values ​​are summed and divided by the total number of pixels in the feature matrix to obtain the average gradient value. .

[0048] A set of feature balancing control parameters is embedded in the first flash memory chip inside the feature estimation unit. This set of parameters includes the reference noise smoothing constant. Its value is 0.005; the first regulating smoothing constant. Its value is 1.25; weight parameter Its value is 0.42; the second anti-interference damping coefficient Its value is 0.35; microconstant Its value is 0.001. In terms of data processing timing, the arithmetic logic unit first calls the divider to calculate the maximum gradient value. With average gradient value Added differential constant The quotient is then compared with the first control smoothing constant. and relative structural complexity factor Perform a series of multiplications and subtract the product from the number 1 as a negative feedback variable; subsequently, the multiplier will factor the relative structural complexity. With the second anti-interference damping coefficient Multiply them together, and combine the product with the minimum gradient value. With weight parameters The product of these factors is added to the aforementioned subtraction result, and the intrinsic scale scaling factor is calculated at the software instruction level. Ungated continuous regression solution.

[0049] Spatial transformation unit receives intrinsic scale scaling factor The feature matrix is ​​written into the asymmetric scaling tensor register of the 2D spatial affine transformation matrix. The reconstruction transformation operator initiates inverse spatial bilinear interpolation resampling on the original feature matrix based on the updated 2D spatial affine transformation matrix. This eliminates the geometric distortion caused by stretching under non-orthophoto perspective from the physical matrix structure and outputs a feature normalization matrix with consistent spatial specifications. The single-scale lightweight classification network consists of a one-dimensional input layer, three consecutively cascaded feature extraction backbone blocks, a global feature transfer layer, and a classification output layer. Each feature extraction backbone block includes a two-dimensional convolutional layer composed of 3x3 convolutional kernels, a batch normalization layer, and a max pooling layer. The channel dimensions of the three feature extraction backbone blocks are 32 channels, 64 channels, and 128 channels, respectively. The feature normalization matrix is ​​imported from the one-dimensional input layer. After continuous channel boosting and spatial downsampling of the three feature extraction backbone blocks, a 128-channel spatial feature tensor is generated at the output of the third feature extraction backbone block.

[0050] Within the global feature transfer layer, the cascaded feature mapping operator opens two parallel fully reduced data paths. The first path calls the averaging unit to perform spatial global averaging compression on each channel of the spatial feature tensor, outputting a 128-dimensional average energy feature vector. The second path calls the maximum extractor to extract the global maximum value for each channel of the spatial feature tensor, outputting a 128-dimensional peak abrupt texture feature vector. The cascaded feature mapping operator combines the average energy feature vector and the peak abrupt texture feature vector into a 256-dimensional intrinsic classification feature vector through the vector concatenation unit.

[0051] The fully connected layer at the output end calls the multiply-accumulate operation array to perform matrix multiplication between the intrinsic classification feature vector and the fixed weight matrix, mapping it to feature scores corresponding to different land surface categories. The classification mapping operator then calls the exponential transformation unit to perform normalized probability calculation on the feature scores, outputting the target category label and decision confidence probability value. The control command is converted into the corresponding land surface target category control protocol based on the label and probability value, enabling the UAV image automatic classification system to maintain the continuous response of classification label output under external interference such as local strong backlight and altitude fluctuations, and reducing the impact of feature energy collapse on affine alignment stability.

[0052] Example 4: This example combines Figures 1 to 2 This paper describes an automatic classification method for space-based observation images, such as... Figure 1 As shown, the original airborne observation image is acquired and converted into the original feature matrix. The local contrast and gray-level variance are calculated. When the local contrast is lower than 20% of the preset global baseline variance, the first control smoothing constant corresponding to the sliding window region is increased proportionally to achieve step S1. Combined with the increased first control smoothing constant, the spectral density of the feature matrix change of the original feature matrix is ​​extracted using the second-order difference operator. At the same time, the anisotropic operator is called to obtain the bidirectional gradient histogram components. An asymmetric scaling tensor is constructed based on the ratio of the bidirectional gradient histogram components to achieve step S2. The intrinsic scale scaling coefficient for global smooth convergence is derived using the spectral density of the feature matrix change. The shear deformation rate of the two-dimensional affine transformation matrix in different coordinate axis directions is corrected based on the asymmetric scaling tensor to achieve step S3. A two-dimensional affine transformation matrix is ​​constructed from the intrinsic scale scaling coefficient and the corrected shear deformation rate. The original feature matrix is ​​resampled to output a feature normalization matrix with consistent specifications. Pooling and vector mapping are performed on the feature normalization matrix in a single-scale space. The final decision confidence and classification result are output to achieve step S4.

[0053] like Figure 2As shown, after system initialization and startup, it enters static calibration mode to establish the initial geometric correspondence between the feature normalization matrix and the original feature matrix. After receiving the self-test pass signal, it switches to standard operating condition to focus on the geometric scale scaling factor solution term to complete cross-space geometric deformation alignment. When the local contrast is lower than 20% of the preset global reference variance, it switches from standard operating condition to meteorological degradation operating condition to continuously adjust the abnormal divergence of the intrinsic scale scaling factor limit coefficient. When the local contrast recovers to the standard range, it returns from meteorological degradation operating condition to standard operating condition. In standard operating condition or meteorological degradation operating condition, when the channel activation variance exceeds the preset stability critical threshold and continues for more than five cycles, it transitions to transient unstable state to make the soft scaling factor exponentially decay and adjust the values ​​of each dimension of the intrinsic classification feature vector. When the channel activation variance recovers to below the critical threshold, it returns from transient unstable state to standard operating condition.

[0054] Example 5: In this example, before the image classification system is deployed in the field, an initial measurement benchmark for the hardware system is established through a pre-baseline calibration procedure. Before the UAV airborne observation payload first flies over a pre-set specific industrial operation area, the optical imaging payload is aligned with a ground standard calibration blackbody with known reflectivity. The digital signal processor reads the original airborne observation image of the current frame and calculates the global grayscale variance. When the global grayscale variance Deviation from the reference thermal noise threshold solidified in the memory cell At that time, the arithmetic logic unit executes the calibration operator to calculate the hardware gain compensation factor. Its calculation formula is expressed as follows ,in, This is the hardware gain compensation factor. The reference thermal noise threshold for curing. To measure the global grayscale variance, This is a microconstant with a value of 0.001, used to prevent denominator collapse, and is the calculated hardware gain compensation factor. The multiplier register of the hardware input interface is directly written to perform amplitude correction on the raw airborne observation image input later, eliminating the static gain drift caused by long-term operation of the imaging sensor, so that the range between the gradient maxima and gradient minima extracted later can reflect the intrinsic gray-scale abrupt change of the surface target.

[0055] The pre-baseline calibration procedure then performs a zeroing operation on the asymmetric scaling tensor register of the two-dimensional affine transformation matrix inside the digital signal processor, and resets the intrinsic scale scaling coefficients. The default value is locked to 1, thus establishing the initial geometric correspondence between the feature normalization matrix and the original feature matrix in the initial operating state of the system. At this time, the UAV performs 10 cycles of smooth gliding maneuvers over the preset industrial operation area. The graphics computing chip collects the intrinsic classification feature vectors continuously output during the gliding cycle and sends them to the sliding time window buffer. The computing unit monitors the change rate of feature mutual information entropy between 10 consecutive sampling points in real time. When the change rate of mutual information entropy falls below the preset convergence criterion of 0.02, the hardware interrupt controller sends a self-test pass signal to the main processor, releasing the flow lock state of the feature mapping operator of the subsequent single-scale lightweight classification network, so that the classification system transitions from static calibration mode to high dynamic online real-time classification operation state, maintaining the numerical stability of the classification label output and the traceability of the decision basis under complex weather conditions.

[0056] Example 6: In this example, before the image processing system is put into production operation, a full-process feature consistency calibration procedure is first executed on an offline calibration platform. The offline calibration platform is equipped with a high-precision area array sensor with optical axis characteristics consistent with the UAV's optical payload, and is configured with a controlled ambient light source and a ground texture simulation board. Both are arranged in a sealed darkroom. After the initialization process is started, the arithmetic logic unit retrieves the stored data from the first memory. A set of predefined reference feature blocks, wherein The value is set to 512; this preset reference feature block set contains 512 discretized reference vectors pre-acquired under standard lighting conditions, covering the pixel distribution features of typical target categories on the ground.

[0057] The arithmetic logic unit reads the sampling potential offset of the current physical hardware interface. and gain deviation coefficient and with a preset compensation model Generate a feature offset compensation vector matrix for the current hardware node. ,in, The hardware feature offset compensation vector matrix is ​​configured with 256 dimensions multiplied by 128 dimensions; This is the sampling potential offset. This is the gain deviation coefficient. This is a nonlinear regression function stored in the fixed code segment, used to establish the mapping relationship from potential offset and gain deviation to feature space distortion. The system compensates for this with a feature offset vector matrix. Preprocessing and correction are performed on the feature vectors of all subsequent input space-based observation images to eliminate static feature drift caused by the thermal noise characteristics of hardware circuit components, so that the feature matrices output by each physical observation node are in a uniform statistical distribution range.

[0058] Nonlinear regression function In its implementation, it consists of a three-layer feedforward multilayer perceptron network structure. The input layer of this network structure contains two nodes, which receive the measured sampling potential offsets. and gain deviation coefficient The hidden layer is configured with 32 channels, and the activation function is the hyperbolic tangent function. The output layer contains 32,768 nodes, which are flattened in software to form a 256-row by 128-column two-dimensional matrix. The values ​​of each output node correspond to the feature offset compensation vector matrix. The corresponding elements in the network, the weight parameters of each connection node are determined in offline state by supervised training of the standard input response curve using the least squares fitting algorithm, and burned into the flash memory, so that the system can output the corresponding compensation matrix according to the underlying hardware potential characteristics.

[0059] After preprocessing and correction, the system cyclically executes a real-time stability self-check procedure during dynamic operations. The arithmetic logic unit sets sampling observation points in the inter-layer data stream of the feature extraction backbone block and calculates the channel activation variance of each level of feature tensor in real time. When the channel activation variance Exceeding the preset stability threshold Furthermore, when the continuous sampling period exceeds 5 units, the feature estimation unit determines that the system has entered a transient unstable region caused by abrupt changes in the complex surface environment.

[0060] Within the aforementioned transiently unstable region, the trigger will apply the soft scaling factor of the classification feature mapping operator. Adjusted to ,in, The soft scaling factor for the current classification feature mapping operator. This is the initial soft scaling factor, set to 1. The response coefficient, which characterizes the sensitivity of the feature space, is set to a value of 0.5. Activate the variance for the channel. As a critical threshold for stability, the system uses a soft scaling factor. Exponential suppression is applied to the values ​​of each dimension of the intrinsic classification feature vector to reduce the weight of unstable feature components in the fully connected layer calculation. This ensures that the decision confidence probability value output by the classification network maintains a gradual trend under abrupt changes, preventing frequent jumps in the target category label caused by violent oscillations in the activation state of the underlying feature matrix.

[0061] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0062] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. An automatic classification method for space-based observation images, characterized in that, Includes the following steps: Step S1: Obtain the original space-based observation image from the space-based observation payload and convert it into the original feature matrix. Calculate the local contrast and gray-level variance of the original feature matrix. When the local contrast is lower than 20% of the preset global reference variance, increase the first control smoothing constant corresponding to the sliding window region proportionally. Step S2: Combine the first adjustment smoothing constant after the adjustment, use the second-order difference operator to extract the characteristic matrix change spectral density of the original characteristic matrix, and at the same time call the anisotropic operator to obtain the bidirectional gradient histogram components, and construct an asymmetric scaling tensor based on the ratio of the bidirectional gradient histogram components. Step S3: Use the spectral density of the characteristic matrix to derive the eigenscale scaling factor for global smooth convergence, and correct the shear deformation rate of the two-dimensional affine transformation matrix in different coordinate axis directions based on the asymmetric scaling tensor. Step S4: Construct a two-dimensional affine transformation matrix using the intrinsic scale scaling factor and the corrected shear deformation rate, resample the original feature matrix, output a feature normalization matrix with consistent specifications, perform pooling and vector mapping on the feature normalization matrix in a single-scale space, and output the final decision confidence and classification result.

2. The automatic classification method for space-based observation images according to claim 1, characterized in that, In step S2, the extraction of the feature matrix change spectral density of the original feature matrix using the second-order difference operator includes the following sub-steps: Step S21, calculate the second-order derivative matrix of the original feature matrix in the horizontal and vertical directions using the second-order difference operator; Step S22, perform Gaussian smoothing filtering on the second-order derivative matrix in combination with the up-adjusted first control smoothing constant to obtain the feature matrix change spectral density.

3. The automatic classification method for space-based observation images according to claim 1, characterized in that, In step S2, the process of calling the anisotropic operator to obtain the bidirectional gradient histogram components includes the following sub-steps: Step S23, using the anisotropic operator to calculate the gradient intensity and gradient direction along the horizontal and vertical axes of the original feature matrix; Step S24, statistically analyzing the gradient intensity distribution along different axes, constructing the horizontal gradient histogram components and the vertical gradient histogram components respectively, and combining them to form the bidirectional gradient histogram components.

4. The automatic classification method for space-based observation images according to claim 1, characterized in that, In step S3, correcting the shear deformation rate of the two-dimensional affine transformation matrix in different coordinate axis directions based on the asymmetric scaling tensor includes the following sub-steps: Step S31, calculate the numerical ratio of the horizontal gradient histogram component to the vertical gradient histogram component to determine the perspective distortion under the non-orthographic projection form; Step S32, fill the off-diagonal elements of the asymmetric scaling tensor according to the perspective distortion, and use the off-diagonal elements to correct the shear deformation rate of the two-dimensional affine transformation matrix in different coordinate axis directions to eliminate nonlinear perspective stretching deformation.

5. The automatic classification method for space-based observation images according to claim 1, characterized in that, In step S3, determining the intrinsic scaling factor using the spectral density of the eigenma matrix variation includes the following sub-steps: Step S33, calculate the range between the gradient maxima and gradient minima of the spectral density of the eigenma matrix variation; Step S34, when the high-frequency eigenenergy slips and the range approaches the zero singularity, trigger a monotonically complementary damped convergence constraint mechanism to stabilize the intrinsic scaling factor within the preset smooth convergence stable region, thus blocking the propagation of numerical divergence to the two-dimensional affine transformation matrix.

6. The automatic classification method for space-based observation images according to claim 1, characterized in that, In step S4, the original feature matrix is ​​resampled and a feature normalization matrix with consistent specifications is output, which includes the following sub-steps: Step S41, using the two-dimensional affine transformation matrix constructed by the intrinsic scale scaling coefficient and the corrected shear deformation rate, a two-way geometric mapping relationship is established from the normalized single-scale space to the original image coordinate space. Step S42: Based on the bidirectional geometric mapping relationship, the original feature matrix is ​​resampled to eliminate the geometric deformation caused by the change in observation height and the change in shooting field of view in the front part of the classification network, and a feature normalization matrix is ​​generated.

7. The automatic classification method for space-based observation images according to claim 1, characterized in that, In step S4, the feature normalization matrix is ​​pooled and vector mapped in the single-scale space, and the final decision confidence and classification result are output. The following sub-steps are included: Step S43, the feature normalization matrix is ​​input into the single-scale lightweight classification network, and the feature normalization matrix is ​​pooled and dimensionality reduced in each channel in the single-scale space to construct the intrinsic classification feature vector. Step S44: Using a fully connected layer and a classification mapping operator, the intrinsic classification feature vector is mapped to the target class probability, and the decision confidence and classification result are output.

8. The automatic classification method for space-based observation images according to claim 7, characterized in that, In step S43, pooling is performed on each channel of the feature normalization matrix in a single-scale space to reduce dimensionality and construct the intrinsic classification feature vector. This includes: jointly using global average pooling and global max pooling to compress the spatial dimension of the feature normalization matrix, extracting the average energy distribution and peak abrupt texture features of each feature channel, and splicing them together to generate the intrinsic classification feature vector.

9. The automatic classification method for space-based observation images according to claim 1, characterized in that, Before calculating the local contrast and grayscale variance of the original feature matrix, a pre-observation parameter correction step is also included: Step S5, acquire the real-time remote sensing altitude data and barometric altitude data of the airborne observation payload, calculate the complementary deviation between the two, and use the complementary deviation to correct the initial spatial resolution benchmark of the original feature matrix, providing a geometric benchmark constraint for determining the intrinsic scale scaling factor.

Images