# US 12,770,391 B1 · Computer implemented system for image capture and neural network training

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 12,770,391 B1
- Document ID: US12770391B1
- Status: Granted
- Kind: B1
- Application number: 19/558,587
- Filed: 2026-03-06
- Granted: 2026-10-20
- Earliest claimed: 2025-11-04
- Assignees: Prince Mohammad Bin Fahd University
- Applicants: Prince Mohammad Bin Fahd University
- Inventors: Muhammad Attique Khan
- Web page: https://patentgrep.com/patent/US12770391B1

## CPC

- G06V 20/10: Terrestrial scenes (scenes under surveillance with static cameras G06V20/52; scenes perceived from the exterior of a vehicle G06V20/56; scenes perceived from the interior of a vehicle G06V20/59)
- G06V 10/56: relating to colour
- G06V 10/764: using classification, e.g. of video objects
- G06V 10/82: using neural networks
- G06V: IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING

## Abstract

A method, and a system for training super-resolution deep neural network to classify coastal area in Remote Sensing (RS) image includes retrieving RS image having first resolution. Further, the method includes processing RS image through multi-block residual attention based Very Deep Super Resolution (VDSR) network to obtain residual image. The method includes generating high-resolution output image by adding residual image to RS image. Further, the method includes generating color feature map using high-resolution output image. The method includes generating resultant feature map by processing high-resolution output image through multimodal inception attention network (M2IAN) including plurality of inception attention processes. Further, the method includes combining the color feature map and resultant feature map. The method further includes learning to classify composite feature image into plurality of coastal area categories. The method includes displaying the composite feature image and classification category.

## Claims

### Claim 1

A system for image capture and neural network training to classify a coastal area in a remote sensing (RS) image, comprising:
an aerial sensor for capturing the RS image, wherein the aerial sensor is mounted on an unmanned aerial vehicle;
a memory storing the RS image having a first resolution;
a machine learning engine configured to
process the RS image through a multi-block residual attention-based Very Deep Super Resolution (VDSR) network to obtain a residual image,
dynamically adjust weights in each block of the multi-block residual attention-based VDSR network having a series of convolutional layers with a residual connection and an attention mechanism, to enhance high-frequency details of the RS image,
generate a high-resolution output image by adding the residual image to the RS image,
generate a color feature map using the high-resolution output image,
generate a resultant feature map by processing the high-resolution output image through a multimodal inception attention network (M2IAN) including a plurality of inception attention processes,
identify special and channel features with each inception attention process of the plurality of inception attention processes including a spatial attention branch and a channel attention branch, respectively,
combine the color feature map and the resultant feature map, including a depth-wise concatenation between the color feature map and the resultant feature map, to generate a composite feature image that captures subtle color variations in coastal areas, and
learn to classify the composite feature image into a plurality of coastal area categories; and
display device to display the composite feature image and classification category.

### Claim 2 (depends on 1)

The system of claim 1, wherein the machine learning engine is further configured to
generate, by a first convolutional layer, an initial feature map from the RS image,
process the initial feature map through sequential residual attention blocks, and
generate, by a final convolutional layer, a residual image.

### Claim 3 (depends on 2)

The system of claim 2, wherein the machine learning engine is further configured to generate the high-resolution output image by adding the residual image to the RS image.

### Claim 4 (depends on 2)

The system of claim 2, wherein each residual attention block is configured to
generate, by a first rectified linear (ReLU) activation function, an output feature map from the initial feature map;
generate, by a series of convolutional layers, a processed feature map from the output feature map;
generate an attention map using a sigmoid activation function;
perform element-wise multiplication between the attention map and the processed feature map to generate a post-processed feature map; and
add the generated post-processed feature map to the initial feature map.

### Claim 5 (depends on 1)

The system of claim 1, wherein the M2IAN is configured to process the high-resolution output image, including
extract color features from the high-resolution output image;
extract preliminary features by passing the high-resolution output image through a plurality of inception attention module blocks; and
concatenate a color feature vector with the preliminary features extracted from the inception attention modules to determine a final feature map.

### Claim 6 (depends on 5)

The system of claim 5, the machine learning engine further configured to
extract color features from the high-resolution output image by
converting RGB color space of the high-resolution output image to HSV color space;
converting the RGB color space to LAB color space; and
calculating statistical features including mean, variance, standard deviation, skewness, kurtosis, harmonic mean, median, and mode for each color channel in the RGB, HSV, and LAB color spaces to determine the color feature vector.

### Claim 7 (depends on 5)

The system of claim 5, wherein each inception attention module block is configured to
identify, by a spatial attention branch comprising a plurality of parallel convolutional paths with different filter sizes, multi-scale spatial features;
generate channel features by a channel attention branch comprising global average pooling and fully connected layers; and
integrate the multi-scale spatial features and the channel features using addition and sigmoid activation.

### Claim 8 (depends on 1)

The system of claim 1, wherein the machine learning engine is further configured to
optimize hyperparameters of the multi-block residual attention-based VDSR network and the multimodal inception-attention network (M2IAN) using Red Fox Optimization (RFO), the RFO configured to:
generate a plurality of initial hyperparameter sets, each set comprising values for epochs, learning rate, batch size, and activation functions;
evaluate each hyperparameter set of the plurality of initial hyperparameter sets using a fitness function based on validation loss;
update hyperparameter values; and
select the hyperparameter set with the best fitness value for arranging the multi-block residual attention-based VDSR network and the multimodal inception-attention network (M2IAN).

### Claim 9 (depends on 1)

The system of claim 1, wherein the plurality of coastal area categories includes at least one from the group consisting of island, beach, wetland, lake, landslide, anchor, river, snowberg, harbor, and harbor and port.

### Claim 10 (depends on 2)

The system of claim 2, wherein the multi-block residual attention-based VDSR network includes 5 residual attention blocks.
