Beach garbage treatment method and system combining bionic sand filtration and AI vision

Abstract

The invention discloses a beach garbage treatment method and system combining bionic sand filtration and AI vision, and relates to the technical field of beach garbage treatment, and the method comprises the steps: obtaining a multispectral image data set, and constructing a dynamic garbage image sequence; constructing a deep learning parallel channel, and determining garbage category information; carrying out twinborn control analysis, generating an initial garbage treatment result, and determining a plurality of interference factors; and based on the multiple interference factors, the initial garbage treatment result is traced back to a deep learning parallel channel for optimization treatment, and the garbage category information is updated in a closed-loop mode to control a sorting execution mechanism to conduct intelligent garbage treatment on the target beach. The technical problems that in the prior art, beach garbage recognition is easily interfered by the environment, and due to the fact that sorting path planning is lack of precise simulation support, recognition precision is low, sorting efficiency is insufficient, and the intelligent degree is insufficient are solved, and the technical effect of improving beach garbage recognition precision, sorting efficiency and the processing intelligent level is achieved.

Description

Technical Field

[0001] This invention relates to the field of beach litter treatment technology, and more specifically to a beach litter treatment method and system that combines biomimetic sand filtration with AI vision. Background Technology

[0002] Current beach litter management faces numerous technological bottlenecks. Traditional methods rely heavily on manual sorting or simple mechanical screening, which is not only inefficient and labor-intensive but also struggles to accurately distinguish different types of litter and cope with the complexities of the beach environment. Existing vision-based litter recognition technologies are often affected by factors such as lighting, sand background, and the changing shape of litter, resulting in insufficient recognition accuracy, weak ability to capture dynamic litter, and a lack of precise simulation planning and closed-loop optimization mechanisms for the sorting process. This leads to unreasonable sorting paths and easily disrupted processing flows, making it impossible to achieve efficient, accurate, and intelligent treatment of beach litter.

[0003] Existing technologies for beach litter identification suffer from technical problems such as low identification accuracy, insufficient sorting efficiency, and inadequate intelligence, which are easily affected by environmental interference and lack precise simulation support for sorting path planning. Summary of the Invention

[0004] This application provides a method and system for beach litter treatment that combines biomimetic sand filtration with AI vision, in order to address the technical problems in existing technologies, such as the susceptibility of beach litter identification to environmental interference, the lack of accurate simulation support for sorting path planning leading to low identification accuracy, insufficient sorting efficiency, and a lack of intelligence.

[0005] In view of the above problems, this application provides a method and system for beach litter treatment that combines biomimetic sand filtration with AI vision.

[0006] The first aspect of this application provides a method for beach litter treatment that combines biomimetic sand filtration with AI vision, the method comprising:

[0007] AI vision is used to acquire multispectral images of the sand on the target beach, resulting in a multispectral image dataset. The beach is then subjected to biomimetic sand filtration based on AI vision to construct a dynamic waste image sequence. A deep learning parallel channel is built to synchronize the multispectral image dataset and the dynamic waste image sequence to the deep learning parallel channel for fusion and recognition, determining waste category information. This waste category information is transmitted to the sorting execution mechanism for twin control analysis, generating initial waste treatment results. Interference analysis is performed based on these initial results to identify multiple interfering factors. Based on these multiple interfering factors, the initial waste treatment results are backtracked to the deep learning parallel channel for optimization, and the waste category information is updated in a closed loop to control the sorting execution mechanism for intelligent waste treatment of the target beach.

[0008] A second aspect of this application provides a beach litter treatment system combining biomimetic sand filtration and AI vision, the system comprising:

[0009] The image sequence construction module is used to acquire multispectral images of the sand on the target beach using AI vision, obtain a multispectral image dataset, and perform biomimetic sand filtration on the target beach using AI vision to construct a dynamic garbage image sequence. The garbage category information determination module is used to construct a deep learning parallel channel, synchronize the multispectral image dataset and the dynamic garbage image sequence to the deep learning parallel channel for fusion recognition, and determine the garbage category information. The interference factor determination module is used to transmit the garbage category information to the sorting execution mechanism for twin control analysis, generate an initial garbage treatment result, and perform interference analysis based on the initial garbage treatment result to determine multiple interference factors. The garbage treatment module is used to backtrack the initial garbage treatment result to the deep learning parallel channel for optimization based on the multiple interference factors, and update the garbage category information in a closed loop to control the sorting execution mechanism to perform intelligent garbage treatment on the target beach.

[0010] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0011] AI-based vision is used to acquire multispectral images of the sand on the target beach, resulting in a multispectral image dataset. The beach is then subjected to biomimetic sand filtration based on AI vision to construct a dynamic sequence of waste images. A deep learning parallel channel is built for fusion recognition to determine waste category information. This waste category information is transmitted to the sorting execution mechanism for twin control analysis, generating initial waste processing results and identifying multiple interfering factors. The initial waste processing results are then backtracked to the deep learning parallel channel for optimization, and the waste category information is updated in a closed loop to control the sorting execution mechanism for intelligent waste processing on the target beach. This achieves the technical effect of improving the accuracy of beach waste identification, sorting efficiency, and the level of intelligent processing. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 A schematic diagram of a beach litter treatment method combining biomimetic sand filtration and AI vision provided in an embodiment of this application;

[0014] Figure 2This is a schematic diagram of a beach litter treatment system combining biomimetic sand filtration and AI vision, provided as an embodiment of this application.

[0015] Figure labeling: Image sequence construction module 10, waste category information determination module 20, interference factor determination module 30, waste processing module 40. Detailed Implementation

[0016] This application provides a method and system for beach litter treatment that combines biomimetic sand filtration with AI vision, in order to address the technical problems in existing technologies, such as the susceptibility of beach litter identification to environmental interference, the lack of accurate simulation support for sorting path planning leading to low identification accuracy, insufficient sorting efficiency, and a lack of intelligence.

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0018] Example 1, as Figure 1 As shown, this application provides a beach litter treatment method combining biomimetic sand filtration and AI vision, the method comprising:

[0019] Step S100: Based on AI vision, multispectral imaging is performed on the sand of the target beach to acquire a multispectral image dataset. Based on AI vision, biomimetic sand filtration is performed on the target beach to construct a dynamic garbage image sequence.

[0020] Specifically, multispectral imaging is conducted on the surface and shallow sand of the target beach to acquire a multispectral image dataset. First, multispectral acquisition and analysis are performed on the area, setting an initial parameter set including multiple acquisition angle parameters and multiple scanning path parameters. Then, the target beach is regionally gridded, and a regional grid map is created. Subsequently, AI vision is activated based on the acquisition angle parameters, and the regional grid map is scanned grid by grid according to the scanning path parameters to generate multi-band grid images. These images are then registered according to the regional grid map and fused pixel by pixel to construct a multispectral image dataset. Next, AI vision traverses and processes this multispectral image dataset to generate sand filtration control commands. These commands activate a biomimetic sand filter device to perform sand filtration analysis and set the screening boundary parameters for the vibrating sand filter belt. The screening area is then delineated based on the screening boundary parameters, and multiple image sensors are deployed within the area. These devices acquire multiple continuous motion images during the biomimetic sand filtration process. Finally, the multiple continuous motion images are 3D stitched together according to the spatial position information of the image sensors to complete the construction of a dynamic waste image sequence.

[0021] Step S200: Construct a deep learning parallel channel to synchronize the multispectral image dataset and the dynamic garbage image sequence to the deep learning parallel channel for fusion recognition and to determine the garbage category information.

[0022] Specifically, a deep learning parallel channel consisting of two sub-channels is constructed. The first deep learning channel adopts a multi-branch convolutional neural network architecture. Its first branch uses three layers of convolutional-pooling units and batch normalization layers to specifically process visible light band images. It accurately captures spatial texture features such as color distribution and contour morphology of waste by using convolutional kernels of different sizes. The second branch adds a spectral attention module to process near-infrared band images, deeply mining the differential response features of sand and waste under the near-infrared spectrum, effectively identifying hidden waste that is difficult to distinguish under visible light. The second deep learning channel adopts a hybrid architecture of three-dimensional convolutional neural network and long short-term memory network. The three-dimensional convolutional neural network performs operations on dynamic waste image sequences through multiple sets of three-dimensional convolutional kernels, simultaneously extracting the spatial morphological features of waste in a single frame image and the inter-frame motion correlation information to generate a complete spatiotemporal feature map. The long short-term memory network uses a gating mechanism to filter key temporal information and filter instantaneous interference such as wave fluctuations and sudden changes in light, accurately modeling the complete motion dependency relationship of waste from separation from sand and movement along the filter belt to discharge. After the dual-channel construction is completed, a deep learning parallel channel is formed. Subsequently, the multispectral image dataset is synchronized to the two branches of the first deep learning channel, extracting spectral dimensional features in the visible light band and near-infrared band respectively. These two types of features are then integrated with the spatial texture features of the multispectral image dataset to construct the first feature vector. Simultaneously, the dynamic waste image sequence is synchronized to the second deep learning channel. Combining the extracted spatiotemporal features and temporal dependencies of the temporal motion, morphological change features and temporal motion features are further extracted to construct the second feature vector. Next, the first and second feature vectors are concatenated to obtain an initial fused feature vector. After dimensionality reduction, a fused feature vector is generated. Multimodal dynamic calibration is then performed on the fused feature vector to obtain multimodal feature correlation coefficients. Based on these coefficients, the fused feature vector is input into a classifier to calculate classification confidence, resulting in multiple confidence scores. Based on these scores, classification regression is performed on the fused feature vector to determine the spatial location information of the waste. Finally, the spatial location information and confidence scores are combined to complete waste classification. This not only clarifies the category of conventional waste but also identifies and labels biological interference information, ultimately forming complete waste category information including biological interference information.

[0023] Step S300: The waste category information is transmitted to the sorting execution mechanism for twin control analysis to generate initial waste treatment results. Based on the initial waste treatment results, interference analysis is performed to identify multiple interference factors.

[0024] Specifically, a digital twin control platform is built corresponding to the sorting execution mechanism. This platform uses 3D modeling technology to replicate the mechanical structure, transmission system, and operation process of the execution mechanism at a 1:1 scale. It also integrates real-time updated virtual environmental parameters such as beach sand hardness, tide level changes, and wind force, forming a simulation space highly similar to the real operation scenario. After transmitting complete waste category information, including biological interference information, waste type, spatial location, and confidence score, to the platform, the platform will conduct intelligent simulation sorting analysis based on waste distribution characteristics and environmental parameters. Using path planning algorithms, it will generate multiple candidate sorting paths covering different operation priorities and adapted to different waste types. Each path includes the action sequence of the execution mechanism and its operation... The process involves several steps: First, a time-series and obstacle avoidance strategy are implemented. Then, dynamic simulations are performed on each candidate sorting path to accurately simulate the motion posture, grasping force, and waste separation effect of the actuator under different sand resistance and wind interference conditions. The simulation results output include multi-dimensional indicators such as sorting success rate, operation time, energy consumption, and equipment wear. These simulation results are then filtered using a multi-objective optimization algorithm, with the optimal target sorting path retrieved as an index and incorporated into the initial waste processing results. Finally, based on the initial waste processing results, interference analysis is conducted. Combining the actual working conditions such as the beach environment, sand flow state, and biological activity areas, multiple interference factors affecting waste sorting accuracy and efficiency are identified and determined, providing data support for subsequent closed-loop optimization.

[0025] Step S400: Based on the multiple interference factors, the initial waste treatment result is backtracked to the deep learning parallel channel for optimization processing, and the waste category information is updated in a closed loop to control the sorting and execution mechanism to carry out intelligent waste treatment on the target beach.

[0026] Specifically, the process involves analyzing the impact of multiple identified interference factors on waste sorting, clarifying the interference coefficients for each factor, and then mapping these interference factors to a multispectral image dataset. First, environmental annotation is performed on the multispectral image dataset to determine environmental metadata. Then, the environmental metadata is combined with the interference coefficients to match the data blocks to be analyzed. By setting a screening threshold, target data blocks exceeding the threshold are extracted and aggregated according to the interference factors, generating the first set of filtered data. Simultaneously, the interference factors are mapped to a dynamic waste image sequence to complete the screening, resulting in the second set of filtered data. Next, data augmentation is performed on the first and second set of filtered data to construct an incremental optimization dataset. Then, the initial waste processing results are backtracked to a deep learning parallel channel, and incremental optimization is performed using the incremental optimization dataset as the target. Based on the incremental optimization data, the waste category information is updated in a closed loop, generating updated waste category information. Finally, the sorting execution mechanism is controlled according to the updated waste category information to achieve intelligent waste processing on the target beach.

[0027] In one possible implementation, step S100 further includes:

[0028] Step S110: Perform multispectral acquisition and analysis on the target beach, and set the initial acquisition parameter set, which includes multiple acquisition angle parameters and multiple scanning path parameters.

[0029] Step S120: Traverse the target beach and perform regional gridding to divide the region into a grid map.

[0030] Step S130: Based on the multiple acquisition angle parameters, activate AI vision to scan the area grid map grid by grid according to the multiple scanning path parameters to generate a multi-band grid image.

[0031] Step S140: Register the multi-band grid image according to the regional grid map, perform pixel fusion based on the registration result, and construct the multispectral image dataset.

[0032] Specifically, for the target beach sand, a comprehensive multispectral acquisition analysis is conducted covering both the beach surface and shallow sand. This analysis considers the beach's topography, sand distribution, lighting conditions, and the estimated potential litter distribution areas to develop a reasonable initial parameter set. This initial parameter set includes multiple differentiated acquisition angle parameters, covering different perspectives such as low-altitude vertical and oblique side scans, ensuring the capture of spectral information from different layers of the sand. It also incorporates multiple scanning path parameters, including lateral scanning paths along the coastline, longitudinal scanning paths perpendicular to the coastline, and gridded scanning paths for key areas, thus ensuring comprehensive, seamless multispectral acquisition of the entire target beach.

[0033] Based on the actual geographical area, topographic features, and preliminary assessment information from multispectral data acquisition and analysis of the target beach, the basic rules for grid division were determined, including the size, shape, and regional division priority of grid units. Subsequently, a full-area traversal of the entire target beach was conducted. Following the established rules, different areas of the beach, from the coastline to the nearshore sand boundary and from the higher points of the beach to the intertidal zone, were divided into uniformly sized, non-overlapping, and comprehensive grid units. Each grid unit was assigned a unique spatial coordinate identifier, ultimately forming a regional grid map that accurately locates each area of ​​the beach. This provides a clear spatial division reference for subsequent AI-based multispectral scanning, ensuring the orderly and comprehensive nature of the data acquisition work.

[0034] First, multiple preset acquisition angle parameters are retrieved to activate the AI ​​vision acquisition device and adjust its shooting angle, enabling it to acquire data in multiple dimensions, such as vertical overhead shots and oblique side scans. Then, according to multiple preset scanning path parameters, a scanning path covering the entire target beach area is planned, including different paths such as horizontal along the coastline, vertical along the coastline, and densification in key areas. Subsequently, the AI ​​vision acquisition device will perform a grid-by-grid scan of the divided area map according to the planned scanning path. For each grid unit, multi-band imaging is performed from different preset acquisition angles to acquire image data in different bands such as visible light and near-infrared. Finally, the multi-band images corresponding to each grid unit are classified and integrated to form a multi-band grid image that can accurately correspond to each grid area of ​​the beach, laying the data foundation for subsequent image registration and fusion.

[0035] Using a segmented regional grid map as a spatial reference, precise image registration is performed on the generated multi-band grid images. Based on the unique spatial coordinates of each grid cell, multi-band images of the same grid cell acquired from different acquisition angles and scanning paths are spatially aligned to eliminate image misalignment caused by device shooting angle deviations and path displacements, ensuring a one-to-one correspondence of pixel positions in each band image within the same grid. After image registration, pixel-level fusion processing is performed on the registered multi-band images of each grid cell, integrating spectral information from different bands such as visible light and near-infrared into the same image data layer. This preserves the unique spectral characteristics of each band while achieving complementarity of multi-dimensional information. Finally, the fused images of all grid cells are integrated according to the spatial distribution of the regional grid, constructing a multispectral image dataset covering the entire target beach area, possessing both multi-band spectral features and precise spatial positioning.

[0036] In one possible implementation, step S100 further includes:

[0037] Step S150: Process the multispectral image dataset based on AI vision to generate sand filter control instructions.

[0038] Step S160: Activate the bionic sand filter device to perform sand filtration analysis by means of the sand filter control command, and set the screening boundary parameters of the vibrating sand filter belt.

[0039] Step S170: Delineate the screening area based on the screening boundary parameters, and deploy multiple image sensing devices in the screening area.

[0040] Step S180: Biomimetic sand filtration and sieving are performed using the multiple image sensing devices to obtain multiple continuous motion images.

[0041] Step S190: The multiple continuous motion images are stitched together in three dimensions according to the spatial position information of the multiple image sensing devices to construct the dynamic garbage image sequence.

[0042] Specifically, the AI ​​vision algorithm framework invokes a multispectral image feature analysis model to perform a full traversal of the constructed multispectral image dataset. Pre-defined spectral feature recognition operators are used to extract the spectral differences between sand and potential waste within each grid cell. Simultaneously, a texture recognition module analyzes the spatial texture information of the image, distinguishing between uniform sand texture and irregular waste texture. A target detection algorithm then locates suspected waste areas, marking their spatial coordinates, area, and particle size range. This analytical data is then input into a sand filter parameter decision model. This model, combined with environmental auxiliary data such as beach sand looseness and tide level, calculates the optimal operating area of ​​the biomimetic sand filter device, the start and stop times of the vibrating sand filter belt, and the vibration frequency, among other core control parameters. Finally, these parameters are encapsulated into standardized sand filter control commands using an industrial control communication protocol. These commands include device start signals, zoning operation instructions, and screening intensity adjustment parameters, thereby achieving precise control over subsequent biomimetic sand filter operations.

[0043] The generated sand filter control commands are transmitted to the main control module of the biomimetic sand filter device via an industrial bus. After verifying the integrity of the commands, the main control module activates the device to enter the sand filtration operation preparation state. The device's built-in sand filtration analysis program then starts, combining information such as the distribution range of garbage and the particle size distribution of sand identified from multispectral image data to quantitatively assess the screenability of the beach sand. Simultaneously, based on the boundary contour between garbage and sand, the device's operating radius, and other actual operating conditions, precise screening boundary parameters are set for the vibrating sand filter belt. These parameters specifically include the effective lateral boundary of the sand filter belt, the longitudinal screening depth boundary, and the gradient change boundary of the screening pore size. The lateral boundary matches the width of the beach garbage distribution along the shore, the longitudinal depth boundary corresponds to the burial depth of garbage in the shallow sand, and the pore size gradient boundary is set differently according to the garbage particle size in different areas. This ensures that the vibrating sand filter belt can efficiently screen garbage from the sand while avoiding over-screening that could damage the sand structure or cause small-sized garbage to be missed, providing reliable parameter support for subsequent precise screening.

[0044] Based on the set screening boundary parameters of the vibrating sand filter belt, combined with the operating range of the biomimetic sand filter device and the actual topography of the beach, the corresponding screening area is precisely delineated. The lateral range of this area matches the effective operating width of the sand filter belt, while the longitudinal range covers the burial depth of shallow sand debris and the movement trajectory range of the sand filter belt, while avoiding non-operating areas such as ecologically sensitive areas of the beach. Subsequently, for the delineated screening area, according to the movement trajectory of the debris during the sand filtration process and the distribution of visual monitoring blind spots, the location of image sensing devices is planned. Image sensing devices are deployed at key locations such as the feed end, middle sorting area, and discharge end of the sand filter belt, as well as at the spatial connection points of the area edges. The spacing of the devices is adapted to their monitoring field of view and resolution to ensure that the monitoring range of each device can achieve coverage without blind spots. At the same time, each device is positioned and marked using preset spatial coordinates, providing a precise positional reference for subsequent 3D image stitching.

[0045] First, all image sensors deployed within the screening area are simultaneously activated, and their frame rate and exposure parameters are adjusted to match the operating speed of the biomimetic sand filter device, ensuring complete capture of the dynamic trajectory of the waste during the sand filtration process. Once the biomimetic sand filter device starts up and enters a stable screening phase, each image sensor continuously captures the entire process of feeding, screening, and discharging within the sand filter belt according to a preset acquisition frequency. The focus is on capturing the displacement changes, morphological flipping, and separation of the waste after the sand is removed, while also recording the position coordinates and contour features of the waste at different screening stages. During this process, the device automatically filters out invalid images caused by sand vibration, retaining only valid frames containing the waste target, ultimately forming multiple sets of continuous motion images ordered along a timeline, corresponding to different monitoring points.

[0046] The precise spatial location information of all image sensing devices within the screening area is retrieved to establish a unified three-dimensional spatial coordinate system. The installation coordinates and shooting angles of each device are then entered into the coordinate system for spatial calibration. Subsequently, multiple sets of continuous motion images are preprocessed to eliminate inconsistencies in image brightness and resolution caused by differences in hardware parameters between different devices. Based on the spatial location of each device, images from different monitoring points at the same time point are spatially aligned, mapping the two-dimensional images to their corresponding spatial locations in the three-dimensional coordinate system. To ensure the temporal continuity of waste movement, continuous image frames from each point are sorted along the timeline, while simultaneously completing the image transitions between adjacent monitoring areas to eliminate image gaps caused by blind spots. Finally, a three-dimensional image fusion algorithm integrates all spatially aligned and temporally sorted continuous motion images to form a dynamic waste image sequence that combines spatial depth and temporal movement. This sequence fully presents the spatial displacement, morphological changes, and other dynamic characteristics of waste during the biomimetic sand filter screening process.

[0047] In one possible implementation, step S200 further includes:

[0048] Constructing the first deep learning channel and the second deep learning channel:

[0049] S1: The first deep learning channel adopts a multi-branch convolutional neural network architecture, which includes a first branch and a second branch. The first branch is configured to process visible light band images, and the second branch is configured to process near-infrared band images.

[0050] S2: The second deep learning channel adopts a hybrid architecture of three-dimensional convolutional neural network and long short-term memory network. The three-dimensional convolutional neural network is used to extract spatiotemporal features from dynamic garbage image sequences, and the long short-term memory network is used to model the temporal dependency of garbage movement.

[0051] The first deep learning channel and the second deep learning channel are analyzed in parallel to construct the deep learning parallel channel.

[0052] Specifically, in the sub-process of constructing the deep learning parallel channel, the first deep learning channel is built. This channel adopts a multi-branch convolutional neural network core architecture, and its first and second branches are equipped with complete convolutional layers, pooling layers, and batch normalization layers to ensure that each branch has independent and efficient feature extraction capabilities. The first branch is specifically configured to process visible light images. Its convolutional layers capture basic visual features such as color distribution and contour morphology in the image through convolutional kernels of different sizes. The pooling layer is responsible for downsampling the feature map to reduce data dimensionality, and the batch normalization layer can standardize the data distribution and improve the model convergence efficiency, thereby achieving accurate analysis of visible light image information. The second branch is specifically responsible for processing near-infrared images. Relying on its internal convolutional layers, it can explore the differential response features of sand and garbage under the near-infrared spectrum. The pooling layer and batch normalization layer simultaneously complete dimensionality reduction and data normalization processing, thereby effectively identifying hidden garbage that is difficult to distinguish under conventional visible light, laying the spectral feature foundation for subsequent multimodal feature fusion.

[0053] In building the second deep learning channel, a hybrid architecture combining a 3D convolutional neural network (3D-CNN) and a long short-term memory (LSTM) network is employed to achieve multi-dimensional feature analysis of dynamic waste image sequences. The 3D-CNN performs convolution operations on the temporal frames and spatial pixel dimensions of the dynamic waste image sequence. This not only extracts spatial morphological features such as the shape and size of waste within a single frame but also captures the displacement, flipping, and separation of waste between adjacent frames, thus integrating spatiotemporal features that combine spatial distribution and temporal variation to fully reconstruct the dynamic trajectory of waste during the biomimetic sand filter screening process. Meanwhile, the LSTM network models the temporal logic of waste movement based on the spatiotemporal features output by the 3D-CNN. Through its unique gating structure, it filters and retains key temporal information while discarding invalid noise, accurately uncovering the temporal dependencies of waste separation from the sand, movement along the filter belt, and eventual discharge, thereby achieving a deep understanding of the dynamic behavior of waste.

[0054] After independently building the first and second deep learning channels, a parallel integration mechanism will be used to collaboratively network the two functionally differentiated channels, thereby constructing a complete parallel deep learning channel. First, a feature interaction and data synchronization link will be established between the channels to ensure that the band feature data processed by the first deep learning channel for multispectral images and the spatiotemporal and temporal feature data parsed by the second deep learning channel for dynamic garbage image sequences can be synchronously transmitted and processed in parallel within the same computing power framework, eliminating the data processing latency difference between the two channels. Simultaneously, a unified feature fusion interface will be configured, reserving a connection port between the first feature vector output by the first deep learning channel and the second feature vector output by the second deep learning channel, providing a data interaction foundation for subsequent multimodal feature fusion. In addition, a channel collaborative scheduling mechanism will be established to intelligently allocate the computing power ratio of the two channels according to the type of input data, such as static multispectral images or dynamic image sequences. When processing multispectral image datasets, the computing power resources of the first deep learning channel will be prioritized. When processing dynamic garbage image sequences, the computing module of the second deep learning channel will be activated simultaneously to achieve efficient collaboration and optimal resource allocation between the two channels. Ultimately, a deep learning parallel channel that can process both static spectral information and dynamic spatiotemporal information will be formed, providing a stable model architecture support for the accurate identification of garbage category information.

[0055] In one possible implementation, step S200 further includes:

[0056] The multispectral image dataset is synchronized to the first branch and the second branch of the first deep learning channel of the deep learning parallel channel to extract the spectral dimension features of the visible light band and the spectral dimension features of the near-infrared band.

[0057] The visible light band spectral dimension features and the near-infrared band spectral dimension features are integrated according to the spatial texture features of the multispectral image dataset to construct a first feature vector.

[0058] The dynamic garbage image sequence is combined with the spatiotemporal features and the temporal dependencies and synchronized to the second deep learning channel of the deep learning parallel channel to extract morphological change features and temporal motion features, and construct a second feature vector.

[0059] The first feature vector and the second feature vector are fused using multimodal feature fusion to generate a fused feature vector. The fused feature vector is then input into a classifier to determine the waste category information.

[0060] Specifically, in the feature extraction stage of waste category identification, the previously constructed multispectral image dataset is first synchronously transmitted to two independent branches of the first deep learning channel in the deep learning parallel channel according to the spectral type. The visible light images from the multispectral image dataset are directed to the first branch, which, relying on its complete convolutional, pooling, and batch normalization layers, performs layer-by-layer feature analysis on the visible light images. Through convolutional operations, it captures key information such as color differences, contour shapes, and surface textures between waste and sand, ultimately extracting visible light spectral dimension features that characterize the target area. Meanwhile, the near-infrared images from the multispectral image dataset are sent to the second branch. The second branch utilizes its dedicated feature processing link to mine the differentiated response features of sand and different types of waste under the near-infrared spectrum, identifying hidden waste targets that are difficult to distinguish under the visible light band, and thus accurately extracting near-infrared spectral dimension features.

[0061] After extracting the spectral features of the visible and near-infrared bands, the two types of features are systematically integrated based on the spatial texture features of the multispectral image dataset to construct the first feature vector. First, the extracted visible spectral features, including color differences and contour morphology, and the near-infrared spectral features, covering infrared spectral response differences and concealed target identification, are spatially aligned. Using the regional grid map of the multispectral image dataset as a benchmark, the two types of spectral features within the same grid cell are precisely matched to ensure a one-to-one correspondence between feature information and the actual spatial location of the beach. Then, the spatial texture features of the multispectral image dataset are retrieved. These features contain key information such as the density and direction of sand and debris textures. This serves as the link for feature integration. A feature weighting algorithm assigns corresponding weights to the spectral features of different regions. For grid cells with significant texture differences and suspected debris concentrations, the weight of their spectral features is increased; for sand areas with uniform texture and no obvious debris, the feature weights are appropriately reduced to decrease redundant information. Finally, the two types of spectral dimension features, which have undergone spatial alignment and weight calibration, are reconstructed and integrated according to the preset vector dimension standard. The scattered feature information is transformed into structured vector data, and finally a first feature vector with both spectral resolution and spatial positioning attributes is formed.

[0062] In the dynamic feature extraction and second feature vector construction stage, the previously generated dynamic waste image sequence, combined with the spatiotemporal features extractable by the 3D convolutional neural network and the temporal dependencies modeled by the long short-term memory network, is synchronously transmitted to the second deep learning channel of the deep learning parallel channel for specialized analysis. First, the 3D convolutional neural network in the channel performs 3D convolution operations on consecutive frames of the dynamic waste image sequence. While capturing the static morphology of waste such as spatial contours and particle size in a single frame image, it also mines the dynamic changes such as the positional shift and orientation flip of waste between adjacent frames, thereby extracting the morphological change features of waste during the screening process. Subsequently, the long short-term memory network, based on its existing temporal dependency modeling capabilities, performs in-depth temporal dimension mining on the feature data output by the 3D convolutional neural network, identifies the complete motion trajectory of waste from sand separation and movement along the filter belt to the discharge, and extracts key information such as its motion rate and displacement direction to form temporal motion features. Finally, dimensional calibration and information integration will be performed on these two types of dynamic features. The spatial attributes of the morphological change features will be associated and matched with the temporal attributes of the temporal motion features. The feature data will be reconstructed according to the preset vector structure, and finally a second feature vector with both dynamic behavior and spatiotemporal correlation attributes will be constructed.

[0063] The first feature vector, which integrates the spectral dimensional features and spatial texture information of the visible / near-infrared bands, is concatenated with the second feature vector, which contains the morphological change features and temporal motion features of the garbage, at the data level to form an initial fused feature vector that covers both static spectral spatial attributes and dynamic spatiotemporal motion attributes. Subsequently, the initial vector is subjected to dimensionality reduction processing, and redundant and invalid feature dimensions are removed by an algorithm to obtain the fused feature vector.

[0064] Next, the fused feature vector undergoes multimodal dynamic calibration, traversing the feature data of different modalities within the vector, calculating the correlation between spectral features and dynamic behavioral features, and obtaining multimodal feature correlation coefficients. Based on these coefficients, the fused feature vector is synchronized to the classifier, which calculates classification confidence for the feature information within the vector, outputting multiple confidence scores for different waste types such as plastic, foam, and shells. Then, based on these confidence scores, classification regression is performed on the fused feature vector to accurately locate the specific spatial position information of each waste item. Finally, the spatial position information of the waste is correlated and matched with the confidence scores to perform waste classification determination, while also identifying special information such as biological interference, ultimately determining complete and accurate waste category information.

[0065] In one possible implementation, step S200 further includes:

[0066] The first feature vector and the second feature vector are concatenated to construct an initial fused feature vector.

[0067] The initial fusion feature vector is used for dimensionality reduction to construct a fusion feature vector.

[0068] Multimodal dynamic calibration is performed by traversing the fused feature vectors to obtain the multimodal feature correlation coefficients.

[0069] The fused feature vector is synchronized to the classifier according to the multimodal feature correlation coefficient to calculate the classification confidence score, thereby obtaining multiple confidence scores.

[0070] Based on the multiple confidence scores, the fused feature vector is classified and regressed to determine the spatial location information of the waste.

[0071] The spatial location information of the waste is combined with the multiple confidence scores to perform waste classification and determine the waste category information.

[0072] Specifically, the first feature vector, which has already been constructed, is retrieved. This vector integrates spectral features in the visible light band and the near-infrared band, and is associated with spatial texture information from a multispectral image dataset, possessing static spectral differences and spatial distribution attributes between waste and sand. Simultaneously, the second feature vector is retrieved, which contains morphological change features and temporal motion features extracted from dynamic waste image sequences, characterizing the dynamic behavior of waste during the biomimetic sand filtration process. Subsequently, the two vectors are sequentially concatenated at the data level according to their dimensional indices, integrating the static spectral spatial feature dimension of the first feature vector with the dynamic spatiotemporal motion feature dimension of the second feature vector into the same data framework. This ensures that the initial fused feature vector possesses both static identification criteria and dynamic behavior representation, laying a multi-dimensional data foundation for subsequent dimensionality reduction and accurate classification.

[0073] The initial fused feature vector undergoes data preprocessing to remove outliers and noisy data, ensuring the purity and validity of the input data. Subsequently, specialized dimensionality reduction algorithms such as Principal Component Analysis (PCA) or Linear Discriminant Analysis (LDA) are introduced to analyze the multi-dimensional feature data within the vector, uncovering the intrinsic relationships between different feature dimensions, and selecting the core feature dimensions that contribute most to waste category identification, while discarding redundant and ineffective feature dimensions. During the dimensionality reduction process, the retention ratio of feature information is strictly controlled. While reducing the vector dimensionality, the key information of the static spectral spatial features and dynamic spatiotemporal motion features in the initial vector is preserved to the maximum extent, avoiding feature loss due to excessive dimensionality reduction. Finally, the dimensionality-reduced core feature data is reintegrated to form a fused feature vector that is dimensionally adapted, information-compact, and possesses multimodal recognition value.

[0074] A feature modality annotation tool is used to classify all feature data within the fused feature vector into modes. Visible light spectral features, near-infrared spectral features, and spatial texture features from the first feature vector are labeled as static spectral spatial modes, while morphological change features and temporal motion features from the second feature vector are labeled as dynamic spatiotemporal motion modes. A modal feature index table is also established to enable rapid feature data retrieval. Subsequently, the Pearson correlation coefficient algorithm and mutual information calculation model are used to traverse each set of feature data within the fused feature vector. First, the co-correlation degree between sub-features within the static spectral spatial mode is calculated. Then, the cross-modal correlation degree between the static spectral spatial mode and the dynamic spatiotemporal motion mode is calculated. Simultaneously, environmental metadata such as beach sand looseness and tide level are integrated, and a weighted correction algorithm is used to adjust the preliminary correlation results for environmental adaptation. Next, a feature weight allocation plugin is used to increase the weight of core recognition features such as waste outline morphology and infrared spectral response according to the preset waste classification recognition priority, while decreasing the weight of interfering features such as sand background texture, thereby optimizing the accuracy of the correlation calculation. Finally, the correlation data within and between all modalities are quantified and normalized using data integration tools to generate multimodal feature correlation coefficients that accurately reflect the correlation strength of each feature. The coefficient results are then stored in the feature correlation database to provide data support for the subsequent confidence calculation of the classifier.

[0075] After obtaining the multimodal feature correlation coefficient, the fused feature vector is precisely synchronized to the classifier based on this coefficient to calculate classification confidence, thereby obtaining multiple confidence scores for different types of waste. First, a feature transmission link is established between the fused feature vector and the classifier. Based on the multimodal feature correlation coefficient, different modal features within the fused feature vector are assigned transmission priorities, prioritizing core features with high correlation, such as waste infrared spectral response features and dynamic morphological change features, to ensure that key identification information participates in the classification calculation first. Subsequently, the classifier calls its built-in waste classification and recognition model, which has been pre-trained on a large number of beach waste samples and is adaptable to the recognition of various waste types, including plastics, foam, and shells, as well as biological interference scenarios. The classifier matches the received fused feature vector with various waste feature templates within the model one by one, while simultaneously weighting the matching results based on the multimodal feature correlation coefficient. For dimensions with high feature correlation, their weight in the matching calculation is increased, thereby enhancing the accuracy of the classification results. During the matching process, the classifier calculates the fit between the fused feature vector and each waste type template in real time, converts the fit value into a standardized confidence score, and finally outputs a set of multiple confidence scores corresponding to different waste categories and special interference information. The higher the score, the greater the probability that the fused feature vector corresponds to this type of waste, providing a quantitative basis for subsequent classification regression and waste location.

[0076] The fused feature vectors are correlated with a previously constructed grid map of the beach area, mapping the feature data within the vectors to the spatial coordinate system corresponding to the grid map. Simultaneously, the geolabelling information of each grid cell is imported to provide a spatial reference for location determination. Subsequently, classification and regression modeling is performed on the fused feature vectors using multiple confidence scores. With the confidence score as the core weight parameter, spatial dimension fitting analysis is conducted on various data representing waste characteristics in the vectors, filtering out feature data with confidence scores higher than a preset threshold and locking their corresponding grid cells. During this process, spatial interpolation algorithms are used to complete the connection areas of feature data between adjacent grid cells, eliminating spatial discontinuities in the feature data and removing interfering data with excessively low confidence scores to avoid bias in location determination. Finally, the feature data fitted by classification and regression is precisely matched with the grid coordinates to clarify the specific grid area, geographical location, and distribution range corresponding to each waste target, ultimately generating waste spatial location information that combines accuracy and spatial correlation.

[0077] After determining the spatial location information of waste, this information is comprehensively matched with multiple confidence scores obtained in the early stages to achieve accurate waste sorting and determine the final waste category information. First, a mapping table is established between the spatial location information of waste and the confidence scores. The spatial coordinates, distribution range, and other location information corresponding to each grid area are bound one-to-one with the confidence scores of various types of waste such as plastic, foam, and shells corresponding to the feature vector of that area, clarifying the waste category matching tendency of different spatial areas. Subsequently, a confidence threshold is set to filter the confidence scores of each area, prioritizing categories with confidence scores higher than the threshold as the core waste type of the corresponding area. For areas with confidence scores close to or below the threshold, the environmental characteristics of their spatial location are considered for auxiliary determination, such as the tendency for plastic waste to appear in near-shore areas and the tendency for shell waste to remain in tidal flat areas, to eliminate interference from non-waste targets such as biological disturbances. Next, the judgment results will undergo spatial aggregation processing, integrating areas of the same waste type that are spatially adjacent to each other to clarify the concentrated distribution areas of various types of waste; at the same time, scattered waste locations will be individually labeled to form a complete spatial-category distribution map of waste. Finally, based on this map, complete waste category information including specific waste categories, precise spatial coordinates, distribution scale, and special interference information will be generated.

[0078] In one possible implementation, step S300 further includes:

[0079] Step S310: Construct a digital twin control platform for the sorting execution mechanism, transmit the waste category information to the digital twin control platform for simulation sorting analysis, and plan multiple candidate sorting paths.

[0080] Step S320: Based on the virtual environment parameters of the digital twin control platform, traverse the multiple candidate sorting paths to perform dynamic simulation and obtain multiple waste simulation processing results.

[0081] Step S330: Perform multi-objective optimization based on the multiple waste simulation processing results, use the optimization results as an index to retrieve the multiple candidate sorting paths, and determine the target sorting path.

[0082] Step S340: Add the target sorting path to the initial waste processing result.

[0083] Specifically, a digital twin control platform for the sorting execution mechanism is first built. This platform replicates the mechanical structure, motion logic, and physical environment of the actual sorting equipment and beach operation scenario in a 1:1 ratio. Then, the determined waste category information, including the spatial location, category attributes, and distribution scale of each type of waste, is completely transmitted to the platform. Based on this information, the platform will conduct simulation sorting analysis and automatically plan multiple candidate sorting paths covering different operation areas and adapting to different sorting methods, combining the morphology, weight, and distribution characteristics of different types of waste. This provides a basic solution for subsequent simulation verification.

[0084] Next, the digital twin control platform retrieves built-in virtual environment parameters, such as the softness of beach sand, wind force level, and equipment operating parameters, and performs dynamic simulations on each of the planned candidate sorting paths. During the simulation, it accurately simulates the complete action process of the actuator grabbing or blowing garbage, and calculates in real time the simulation success rate of each path, i.e. the probability of accurate garbage sorting, the estimated time, i.e. the total time to complete the sorting task of that path, and the energy consumption, i.e. the total energy consumption of the equipment operation. These quantitative data are then integrated into multiple garbage simulation processing results for each path.

[0085] Then, a multi-objective optimization algorithm is launched for multiple waste simulation processing results. With high simulation success rate, low expected time consumption and low energy consumption as the core optimization objectives, the algorithm performs weighted calculation and comprehensive ranking on the three key indicators of each candidate path, selects the optimization result with the best overall performance, and then uses the optimization result as a retrieval index to accurately match the corresponding scheme in the candidate sorting path pool, and finally determines the target sorting path that takes into account both operation efficiency and cost.

[0086] Finally, the selected target sorting paths are integrated with the corresponding simulation parameters and operational requirements, and formally added to the initial waste treatment results generated by the system, providing clear and actionable guidance for the subsequent operations of the actual sorting execution agencies.

[0087] In one possible implementation, step S400 further includes:

[0088] Step S410: Based on the multiple interference factors, conduct a waste sorting impact analysis to determine multiple interference impact coefficients.

[0089] Step S420: Map the multiple interference factors to the multispectral image dataset according to the multiple interference influence coefficients for filtering, and generate the first filtering data.

[0090] Step S430: Map the multiple interference factors to the dynamic garbage image sequence according to the multiple interference influence coefficients for filtering, and generate second filtering data.

[0091] Step S440: Perform data augmentation on the first filtered data and the second filtered data to construct an incrementally optimized dataset.

[0092] Step S450: Based on the incremental optimization dataset as the target, the initial waste treatment result is backtracked to the deep learning parallel channel for incremental optimization processing. The waste category information is updated in a closed loop according to the incremental optimization data, and waste category update information is generated to carry out intelligent waste treatment on the target beach.

[0093] Specifically, in determining the interference impact coefficients, a systematic analysis of the impact of several previously identified interference factors, such as instantaneous wind force on the beach, fluctuations in sand moisture content, vibration amplitude deviation of the biomimetic sand filter device, and acquisition angle offset of the image sensing device, is conducted. First, a quantitative assessment model of interference impact is built, linking each interference factor with core indicators of waste sorting, such as AI visual recognition accuracy, biomimetic sand filter screening efficiency, and the grasping accuracy of the sorting execution mechanism, clarifying the logic of each factor's effect on different core indicators. Then, multiple sets of controlled experiments are conducted in a simulated environment using the controlled variable method to test the impact of a single interference factor at different intensities on the sorting process, while recording the fluctuation data of each indicator. Next, the impact weight of each interference factor is calculated based on the experimental data, and the weights are adjusted according to the probability of occurrence in actual beach operation scenarios. Finally, the impact degree of each interference factor is transformed into multiple standardized interference impact coefficients. The higher the coefficient value, the stronger the interference of the factor on the entire waste sorting process, providing a quantitative judgment benchmark for subsequent dataset selection and model optimization.

[0094] A comprehensive environmental annotation process was conducted on the multispectral image dataset, identifying multiple environmental metadata such as sand type, water content, light intensity, and wind speed for each image region, establishing a precise correlation between the image data and the actual beach working environment. Subsequently, multiple interference factors were combined with their corresponding interference impact coefficients and matched against the annotated environmental metadata in the multispectral image dataset to identify image regions affected by these interference factors, defining them as multiple data blocks to be analyzed. Factors with higher interference impact coefficients were assigned higher matching priority to their corresponding data blocks. Next, a clear interference impact screening threshold was set, and all data blocks to be analyzed were quantitatively evaluated. Only data blocks with interference impact exceeding the threshold were extracted as multiple target data blocks, filtering out redundant image data with low or no interference to ensure the targeted nature of the selected data. Finally, these target data blocks were categorized, aggregated, and format-calibrated according to the specific type of interference factor, such as environmental interference and equipment interference, ultimately generating the first-stage filtered data that accurately reflects the changes in the spectral characteristics of sand and debris under different interference scenarios.

[0095] The dynamic waste image sequence undergoes temporal and spatial preprocessing to mark the screening stage, sensor location, and waste movement state corresponding to each frame in the sequence, establishing an index linking the sequence data with the actual operational scenario. Subsequently, multiple interference factors, such as equipment vibration deviation, environmental wind interference, and changes in sand moisture content, are weighted and sorted according to their respective interference impact coefficients; interference factors with higher coefficients have higher data screening priority. Next, the sorted interference factors are mapped one by one to the dynamic waste image sequence. Using a feature matching algorithm, image frames in the sequence that exhibit waste movement trajectory deviation, morphological feature distortion, or blurred image acquisition due to various interferences are retrieved. These affected frames are marked as data to be screened. Simultaneously, an interference impact threshold is set to remove normal sequence segments with interference impact coefficients below the threshold that have no significant impact on the identification of dynamic waste features. Finally, the marked data to be screened is categorized, and image frames are aggregated according to the type of interference factor, retaining typical data of waste dynamic features under each type of interference. Finally, the classified and aggregated image frame data are integrated temporally and calibrated to generate second-screen data that accurately reflects the movement characteristics of garbage in the interference scenario.

[0096] The first and second selected data sets were standardized in format and aligned in dimensions to eliminate differences in storage format and feature dimensions, establishing a mapping between static and dynamic interference features to form a basic fusion dataset. Subsequently, addressing the limited sample size and insufficient coverage of interference scenarios in the basic fusion dataset, multiple data augmentation operations were initiated: for the multispectral images in the first selected data set, image flipping, brightness adjustment, and noise injection were used to expand the samples, simulating the interference spectral features under different lighting and sandy conditions; for the dynamic garbage image sequences in the second selected data set, time-series frame interpolation, motion trajectory offset simulation, and multi-view stitching were performed to recreate the garbage movement state under different wind forces and equipment vibration interference. Simultaneously, a feature fusion algorithm was used to cross-modal superposition of some static spectral interference features and dynamic motion interference features to generate sample data for composite interference scenarios. Finally, all augmented data underwent quality verification, invalid and distorted samples were removed, and the data was classified, labeled, and structured according to interference type, ultimately constructing an incrementally optimized dataset covering multiple types of interference scenarios and possessing both static and dynamic features.

[0097] The initial waste processing results generated earlier, including target sorting paths, basic waste categories, and spatial distribution information, are completely backtracked to the deep learning parallel channel to initiate the incremental optimization process. The first deep learning channel within the channel calls upon the first selected data from the incremental optimization dataset to fine-tune the weight parameters of its multi-branch convolutional neural network, enhancing the model's ability to identify visible and near-infrared spectral features under interference. The second deep learning channel, relying on the second selected data from the incremental optimization dataset, optimizes the spatiotemporal feature extraction logic of the 3D convolutional neural network and the temporal dependency modeling accuracy of the long short-term memory network, improving the model's adaptability to abnormal waste morphology and movement trajectories caused by interference. After the deep learning parallel channel completes incremental optimization, it outputs incrementally optimized data covering various interference scenarios. Based on this data, the original waste category information is updated in a closed loop: on the one hand, it corrects misclassification data caused by interference factors and supplements newly identified waste location information under interference scenarios; on the other hand, it adds interference adaptation labels to the waste category information of different areas based on the interference impact coefficient, clarifying the sorting priority of various types of waste under interference environments. Ultimately, accurate and interference-resistant waste category update information is generated and synchronized to the sorting execution agency. This guides the equipment to adjust its operating strategy based on the updated information, enabling differentiated and intelligent sorting and processing of different types of waste in different areas of the target beach. This achieves closed-loop iteration and efficiency improvement of the entire beach waste treatment process.

[0098] In one possible implementation, step S420 further includes:

[0099] Step S421: Traverse the multispectral image dataset to perform environmental annotation and determine multiple environmental metadata.

[0100] Step S422: Based on the multiple environmental metadata and the multiple interference influence coefficients, map the multiple interference factors to the multispectral image dataset for matching, and determine multiple data blocks to be analyzed.

[0101] Step S423: Set a filtering threshold, extract data blocks greater than the filtering threshold from the plurality of data blocks to be analyzed, and obtain a plurality of target data blocks.

[0102] Step S424: Classify and aggregate the multiple target data blocks according to the multiple interference factors to generate the first filtered data.

[0103] Specifically, the multispectral image dataset is fully traversed and environmentally labeled. By combining environmental perception algorithms with manual verification, key information such as sand type, sand moisture content, real-time light intensity, nearshore wind force, and tide level is labeled for each grid area in the dataset. These parameters that reflect the beach working environment are integrated into multiple environmental metadata, establishing a precise correlation between image data and the actual beach environment.

[0104] The established environmental metadata is combined with previously generated interference coefficients as a matching criterion to map various interference factors, such as abrupt changes in illumination, abnormal sand moisture content, and wind interference, onto a multispectral image dataset. A feature matching algorithm is used to retrieve image regions in the dataset that match the environmental characteristics of each interference factor. For example, strong wind interference is matched to grid images labeled with high wind speeds, and high moisture content interference is matched to sand images labeled with high moisture content. These successfully matched image regions are then designated as multiple data blocks to be analyzed, with factors having higher interference coefficients receiving higher matching weights and priorities for their corresponding data blocks.

[0105] Based on the previously determined distribution characteristics of interference impact coefficients, the noise level of the multispectral image dataset, and the requirements for interference samples in subsequent model optimization, a reasonable screening threshold was determined using statistical analysis methods such as percentile method and ROC curve analysis. This threshold needs to effectively retain data blocks that are significantly affected by interference and have a substantial impact on the identification of garbage spectral features, while filtering out redundant information with weak interference levels. Subsequently, the identified multiple data blocks to be analyzed were quantitatively evaluated one by one. The actual interference impact level corresponding to each data block was calculated using an algorithm, based on the matching degree between environmental metadata and interference factors, the weighted value of interference impact coefficients, and compared with the preset screening threshold. Finally, only the data blocks to be analyzed with interference impact levels greater than the screening threshold were extracted and integrated into multiple target data blocks to ensure that the selected data accurately focuses on the core interference scenarios.

[0106] Several categories of interference factors identified in the early stages were analyzed, such as environmental interference including sudden changes in light intensity, wind interference, and abnormal sand moisture content, and equipment interference including sensor angle deviation and imaging noise, to establish a clear classification index system for interference factors. Subsequently, the core interference factors corresponding to each target data block were identified one by one. Feature matching algorithms were used to confirm which type(s) of interference factors caused the variations in the spectral characteristics of waste and sand in the data block, and the target data block was classified into the corresponding interference factor category based on the weighting of the interference impact coefficient. For example, target data blocks with abnormal spectral reflectance due to strong light were classified into the light interference category, and target data blocks with image distortion due to equipment sensor angle deviation were classified into the equipment sensor interference category. During the classification process, key information such as environmental metadata and the quantitative value of the interference impact degree for each data block were recorded simultaneously to ensure the traceability of the classification results. Finally, all target data blocks under the same interference factor category are standardized in format, including image resolution, spectral band sorting, and data storage format. Feature dimensions are integrated to remove duplicate and redundant feature information, ultimately generating first-screen data that is clearly divided by interference factor type and has both completeness and relevance. This provides standardized and high-quality static interference samples for subsequent fusion with the second-screen data to construct an incremental optimization dataset.

[0107] Example 2, based on the same inventive concept as the beach litter treatment method combining biomimetic sand filtration and AI vision in the aforementioned examples, such as... Figure 2 As shown, this application provides a beach litter treatment system combining biomimetic sand filtration and AI vision. The system and method embodiments in this application are based on the same inventive concept. The system includes:

[0108] The image sequence construction module 10 is used to perform multispectral imaging acquisition of the sand body of the target beach based on AI vision, obtain a multispectral image dataset, perform biomimetic sand filtration screening of the target beach based on AI vision, and construct a dynamic garbage image sequence.

[0109] The waste category information determination module 20 is used to construct a deep learning parallel channel to synchronize the multispectral image dataset and the dynamic waste image sequence to the deep learning parallel channel for fusion recognition and to determine the waste category information.

[0110] The interference factor determination module 30 is used to transmit the waste category information to the sorting execution mechanism for twin control analysis, generate initial waste treatment results, and perform interference analysis based on the initial waste treatment results to determine multiple interference factors.

[0111] The waste treatment module 40 is used to backtrack the initial waste treatment result to the deep learning parallel channel for optimization based on the multiple interference factors, and to update the waste category information in a closed loop to control the sorting and execution mechanism to carry out intelligent waste treatment on the target beach.

[0112] Furthermore, the system is also used to implement the following functions:

[0113] Multispectral data acquisition and analysis are performed on the target beach. An initial set of acquisition parameters is set, which includes multiple acquisition angle parameters and multiple scanning path parameters. The target beach is traversed and the area is divided into a grid map. Based on the multiple acquisition angle parameters, AI vision is activated to scan the area grid map grid by grid according to the multiple scanning path parameters, generating a multi-band grid image. The multi-band grid image is then registered according to the area grid map, and pixel fusion is performed based on the registration result to construct the multispectral image dataset.

[0114] Furthermore, the system is also used to implement the following functions:

[0115] The multispectral image dataset is processed using AI vision to generate sand filter control commands. These commands activate a biomimetic sand filter device for sand analysis, setting screening boundary parameters for the vibrating sand filter belt. A screening area is defined based on these boundary parameters, and multiple image sensors are deployed within this area. Biomimetic sand filter screening is performed using these image sensors to acquire multiple continuous motion images. These continuous motion images are then stitched together in three dimensions according to the spatial position information of the image sensors to construct the dynamic waste image sequence.

[0116] Furthermore, the system is also used to implement the following functions:

[0117] Constructing a first deep learning channel and a second deep learning channel: S1: The first deep learning channel adopts a multi-branch convolutional neural network architecture, which includes a first branch and a second branch. The first branch is configured to process visible light band images, and the second branch is configured to process near-infrared band images. S2: The second deep learning channel adopts a hybrid architecture of a three-dimensional convolutional neural network and a long short-term memory network. The three-dimensional convolutional neural network is used to extract spatiotemporal features from dynamic garbage image sequences, and the long short-term memory network is used to model the temporal dependencies of garbage movement. The first deep learning channel and the second deep learning channel are analyzed in parallel to construct the deep learning parallel channel.

[0118] Furthermore, the system is also used to implement the following functions:

[0119] The multispectral image dataset is synchronized to the first branch and the second branch of the first deep learning channel of the deep learning parallel channel to extract visible light band spectral dimension features and near-infrared band spectral dimension features. The visible light band spectral dimension features and the near-infrared band spectral dimension features are integrated according to the spatial texture features of the multispectral image dataset to construct a first feature vector. The dynamic garbage image sequence is synchronized to the second deep learning channel of the deep learning parallel channel in combination with the spatiotemporal features and the temporal dependency relationship to extract morphological change features and temporal motion features to construct a second feature vector. The first feature vector and the second feature vector are fused using multimodal features to generate a fused feature vector. The fused feature vector is input into the classifier to determine the garbage category information.

[0120] Furthermore, the system is also used to implement the following functions:

[0121] The first feature vector and the second feature vector are concatenated to construct an initial fused feature vector; dimensionality reduction is performed on the initial fused feature vector to construct a fused feature vector; multimodal dynamic calibration is performed on the fused feature vector to obtain multimodal feature correlation coefficients; the fused feature vector is synchronized to the classifier according to the multimodal feature correlation coefficients to calculate classification confidence and obtain multiple confidence scores; classification regression is performed on the fused feature vector based on the multiple confidence scores to determine the spatial location information of the waste; the spatial location information of the waste is combined with the multiple confidence scores to perform waste classification and determine the waste category information.

[0122] Furthermore, the system is also used to implement the following functions:

[0123] A digital twin control platform for the sorting execution mechanism is constructed. The waste category information is transmitted to the digital twin control platform for simulation sorting analysis, and multiple candidate sorting paths are planned. Based on the virtual environment parameters of the digital twin control platform, the multiple candidate sorting paths are traversed to perform dynamic simulation, and multiple waste simulation processing results are obtained. Multi-objective optimization is performed based on the multiple waste simulation processing results, and the optimization results are used as an index to retrieve the multiple candidate sorting paths to determine the target sorting path. The target sorting path is added to the initial waste processing results.

[0124] Furthermore, the system is also used to implement the following functions:

[0125] Based on the multiple interference factors, an impact analysis on waste sorting is performed to determine multiple interference influence coefficients. These interference factors are then mapped to the multispectral image dataset for filtering according to these coefficients, generating first-filtered data. Similarly, the interference factors are mapped to the dynamic waste image sequence for filtering according to these coefficients, generating second-filtered data. The first-filtered data and the second-filtered data are then augmented to construct an incremental optimization dataset. Based on this incremental optimization dataset, the initial waste processing results are backtracked to a deep learning parallel channel for incremental optimization. The waste category information is then updated in a closed loop based on the incremental optimization data, generating updated waste category information for intelligent waste processing of the target beach.

[0126] Furthermore, the system is also used to implement the following functions:

[0127] The multispectral image dataset is traversed for environmental annotation to determine multiple environmental metadata. Multiple interference factors are mapped to the multispectral image dataset for matching based on the multiple environmental metadata and multiple interference influence coefficients, thus determining multiple data blocks to be analyzed. A filtering threshold is set, and data blocks larger than the filtering threshold within the multiple data blocks to be analyzed are extracted to obtain multiple target data blocks. The multiple target data blocks are then classified and aggregated according to the multiple interference factors to generate the first filtered data.

[0128] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0129] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0130] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and variations fall within the scope of this application and its equivalents, this application intends to include such modifications and variations.

Claims

1. A beach litter treatment method combining biomimetic sand filtration and AI vision, characterized in that, The method includes: AI vision is used to perform multispectral imaging of the sand on the target beach to obtain a multispectral image dataset. Based on AI vision, the target beach is subjected to biomimetic sand filtration and screening to construct a dynamic garbage image sequence. A deep learning parallel channel is constructed to synchronize the multispectral image dataset and the dynamic garbage image sequence to the deep learning parallel channel for fusion recognition and to determine the garbage category information. The waste category information is transmitted to the sorting execution mechanism for twin control analysis to generate initial waste treatment results. Based on the initial waste treatment results, interference analysis is performed to identify multiple interference factors. Based on the aforementioned multiple interfering factors, the initial waste treatment results are backtracked to a deep learning parallel channel for optimization, and the waste category information is updated in a closed loop to control the sorting and execution mechanism to carry out intelligent waste treatment on the target beach.

2. The beach litter treatment method combining biomimetic sand filtration and AI vision as described in claim 1, characterized in that, The method involves using AI vision to perform multispectral imaging of the sand on a target beach to acquire a multispectral image dataset. Multispectral data acquisition and analysis are performed on the target beach, and an initial set of acquisition parameters is set, which includes multiple acquisition angle parameters and multiple scanning path parameters. Traverse the target beach and perform regional grid processing to divide the region into a grid map; Based on the multiple acquisition angle parameters, the AI ​​vision is activated to scan the regional grid map grid by grid according to the multiple scanning path parameters, generating a multi-band grid image. The multi-band grid images are registered according to the regional grid map, and pixel fusion is performed based on the registration results to construct the multispectral image dataset.

3. The beach litter treatment method combining biomimetic sand filtration and AI vision as described in claim 1, characterized in that, Based on AI vision, a biomimetic sand filtration process is used to separate target beaches and construct a dynamic sequence of trash images. The method includes: The multispectral image dataset is processed based on AI vision to generate sand filter control instructions. The bionic sand filter device is activated by the sand filter control command to perform sand filtration analysis and set the screening boundary parameters of the vibrating sand filter belt. The screening area is defined based on the screening boundary parameters, and multiple image sensing devices are deployed in the screening area. Multiple continuous motion images are obtained by biomimetic sand filtration and sieving using the multiple image sensing devices. The multiple continuous moving images are stitched together in three dimensions based on the spatial location information of multiple image sensing devices to construct the dynamic garbage image sequence.

4. The beach litter treatment method combining biomimetic sand filtration and AI vision as described in claim 1, characterized in that, The process of constructing parallel channels in deep learning includes the following methods: Constructing the first deep learning channel and the second deep learning channel: S1: The first deep learning channel adopts a multi-branch convolutional neural network architecture, which includes a first branch and a second branch. The first branch is configured to process visible light band images, and the second branch is configured to process near-infrared band images. S2: The second deep learning channel adopts a hybrid architecture of three-dimensional convolutional neural network and long short-term memory network. The three-dimensional convolutional neural network is used to extract spatiotemporal features from dynamic garbage image sequences, and the long short-term memory network is used to model the temporal dependency of garbage movement. The first deep learning channel and the second deep learning channel are analyzed in parallel to construct the deep learning parallel channel.

5. The beach litter treatment method combining biomimetic sand filtration and AI vision as described in claim 4, characterized in that, The method involves synchronizing the multispectral image dataset and the dynamic garbage image sequence to the deep learning parallel channel for fusion and recognition to determine garbage category information. The multispectral image dataset is synchronized to the first branch and the second branch of the first deep learning channel of the deep learning parallel channel to extract the spectral dimension features of the visible light band and the spectral dimension features of the near-infrared band. The visible light band spectral dimension features and the near-infrared band spectral dimension features are integrated according to the spatial texture features of the multispectral image dataset to construct a first feature vector; The dynamic garbage image sequence is combined with the spatiotemporal features and the temporal dependencies and synchronized to the second deep learning channel of the deep learning parallel channel to extract morphological change features and temporal motion features, and construct a second feature vector. The first feature vector and the second feature vector are fused using multimodal feature fusion to generate a fused feature vector. The fused feature vector is then input into a classifier to determine the waste category information.

6. The beach litter treatment method combining biomimetic sand filtration and AI vision as described in claim 5, characterized in that, The method involves fusing the first feature vector and the second feature vector using multimodal features to generate a fused feature vector, and then inputting the fused feature vector into a classifier to determine waste category information. The first feature vector and the second feature vector are concatenated to construct an initial fused feature vector; Based on the initial fused feature vector, dimensionality reduction processing is performed to construct a fused feature vector; Multimodal dynamic calibration is performed by traversing the fused feature vectors to obtain the multimodal feature correlation coefficients; The fused feature vector is synchronized to the classifier according to the multimodal feature correlation coefficient to calculate the classification confidence score, thereby obtaining multiple confidence scores; Based on the multiple confidence scores, the fused feature vector is classified and regressed to determine the spatial location information of the waste. The spatial location information of the waste is combined with the multiple confidence scores to perform waste classification and determine the waste category information.

7. The beach litter treatment method combining biomimetic sand filtration and AI vision as described in claim 1, characterized in that, The waste category information is transmitted to the sorting execution mechanism for twin control analysis to generate initial waste treatment results. The method includes: A digital twin control platform for the sorting execution mechanism is constructed, and the waste category information is transmitted to the digital twin control platform for simulation sorting analysis to plan multiple candidate sorting paths; Based on the virtual environment parameters of the digital twin control platform, dynamic simulation is performed by traversing the multiple candidate sorting paths to obtain multiple waste simulation processing results; Based on the results of the multiple waste simulation processes, multi-objective optimization is performed, and the optimization results are used as an index to retrieve the multiple candidate sorting paths to determine the target sorting path. Add the target sorting path to the initial waste processing result.

8. The beach litter treatment method combining biomimetic sand filtration and AI vision as described in claim 1, characterized in that, Based on the aforementioned multiple interfering factors, the initial waste disposal results are backtracked to a deep learning parallel channel for optimization. The waste category information is then updated in a closed loop to control the sorting and execution mechanism for intelligent waste disposal on the target beach. The method includes: Based on the aforementioned multiple interference factors, an impact analysis of waste sorting was conducted to determine multiple interference impact coefficients; The multiple interference factors are mapped to the multispectral image dataset according to the multiple interference influence coefficients for filtering, thereby generating the first filtering data; The multiple interference factors are mapped to the dynamic garbage image sequence according to the multiple interference influence coefficients for filtering, thereby generating second filtering data; The first and second filtered data are augmented to construct an incrementally optimized dataset. Based on the incremental optimization dataset as the target, the initial waste treatment results are backtracked to the deep learning parallel channel for incremental optimization. The waste category information is updated in a closed loop according to the incremental optimization data, and waste category update information is generated to carry out intelligent waste treatment on the target beach.

9. The beach litter treatment method combining biomimetic sand filtration and AI vision as described in claim 8, characterized in that, The method involves mapping the multiple interference factors to the multispectral image dataset according to the multiple interference influence coefficients for filtering, generating first filtered data, and the method includes: The multispectral image dataset is traversed to perform environmental annotation and determine multiple environmental metadata. Based on the multiple environmental metadata and the multiple interference influence coefficients, the multiple interference factors are mapped to the multispectral image dataset for matching, thereby determining multiple data blocks to be analyzed. Set a filtering threshold, extract data blocks that are greater than the filtering threshold from the multiple data blocks to be analyzed, and obtain multiple target data blocks; The multiple target data blocks are classified and aggregated according to the multiple interference factors to generate the first filtered data.

10. A beach litter treatment system combining biomimetic sand filtration and AI vision, characterized in that: The system is used to implement the beach litter treatment method combining biomimetic sand filtration and AI vision as described in any one of claims 1-9, the system comprising: The image sequence construction module is used to acquire multispectral images of the sand on the target beach based on AI vision, obtain a multispectral image dataset, perform biomimetic sand filtration on the target beach based on AI vision, and construct a dynamic garbage image sequence. The waste category information determination module is used to construct a deep learning parallel channel, synchronize the multispectral image dataset and the dynamic waste image sequence to the deep learning parallel channel for fusion recognition, and determine the waste category information; The interference factor determination module is used to transmit the waste category information to the sorting execution mechanism for twin control analysis, generate initial waste treatment results, perform interference analysis based on the initial waste treatment results, and determine multiple interference factors. The waste treatment module is used to backtrack the initial waste treatment result to the deep learning parallel channel for optimization based on the multiple interference factors, and to update the waste category information in a closed loop to control the sorting and execution mechanism to carry out intelligent waste treatment on the target beach.

Images