Solid waste monitoring and data management system based on artificial intelligence

Abstract

The invention discloses a solid waste monitoring and data management system based on artificial intelligence, and relates to the technical field of water environment monitoring, the system is composed of an image acquisition module, a data preprocessing module, an encoder feature extraction module, an up-sampling fusion module, a detection head module and a data management module, and the system is based on an RW-YOLOv11 architecture. A C3K2Sc feature extraction unit is introduced into an encoder, and floating garbage edge and texture expression is enhanced through space attention and a dynamic channel reconstruction mechanism; a SurfCAU water surface content awareness enhanced up-sampling module is adopted in the neck network, detail compensation of low-resolution features is achieved, and the multi-scale feature fusion quality is improved; a SurfMSDFHead multi-branch structure is adopted in the detection head, multi-scale feature fusion is realized through dynamic weight, and the positioning precision of overlapped garbage is improved by using a Focaler-IOU interval weighting strategy. The system can realize real-time detection, classification and positioning of the floating garbage in the river channel, and has the advantages of high detection precision, good lightweight degree, strong adaptability to complex water surface scenes and the like.

Description

Artificial Intelligence-Based Solid Waste Monitoring and Data Management System Technical Field

[0001] This invention relates to the field of water environment monitoring technology, and more specifically, to an artificial intelligence-based solid waste monitoring and data management system. Background Technology

[0002] Floating debris in waterways is a significant challenge in urban water environment management. Affected by factors such as water flow speed, wind disturbance, surface reflection, and obstruction by aquatic plants, floating debris exhibits complex shapes, significant size variations, and random distribution, posing considerable difficulties for manual inspections. Traditional manual river patrols are not only costly and inefficient but also subjective, failing to meet the needs for real-time, precise, and routine management of solid waste in waterways. Therefore, developing an intelligent detection technology capable of reliably identifying various types of floating debris in complex aquatic environments has become a crucial requirement in the fields of smart water management and ecological governance.

[0003] With the development of deep learning and computer vision technologies, object detection algorithms have made significant progress in image recognition tasks. Existing methods such as Faster R-CNN, YOLOv5, and YOLOv8, while achieving good detection results in general scenarios, still have many limitations when directly applied to river scenarios. On the one hand, ripples, reflections, and shadows in the river environment introduce a large amount of background noise, making traditional convolutional feature extraction structures prone to false positives and false negatives. On the other hand, floating debris is mostly small-sized targets such as plastic bottles, foam, and packaging bags, which occupy a very small proportion from a distance, making it difficult for conventional detection frameworks to capture their effective features. Furthermore, when multiple targets are stacked, close together, or deformed by water movement on the water surface, the feature branching and boundary regression capabilities of existing detection heads are often insufficient, leading to significant localization errors and affecting detection accuracy.

[0004] To address these issues, related research has gradually attempted to improve lightweight network structures, enhance multi-scale feature representation capabilities, and introduce attention mechanisms to improve feature separation capabilities in complex backgrounds. However, their adaptability in river scenarios remains limited. For example, some methods are only optimized for specific types of waste, resulting in insufficient performance in small target detection; some methods significantly increase model computation while improving accuracy, making them difficult to run in real time on edge computing devices such as unmanned surface vessels; and some methods lack sufficient multi-scale feature fusion, failing to effectively handle the problem of large differences in the size and uneven distribution of floating waste on the river surface. Therefore, there is an urgent need for an intelligent waste monitoring system that addresses the complex water surface characteristics of rivers, possesses small target enhancement capabilities, supports dynamic fusion of multi-scale features, and can run stably and in real time on embedded platforms to improve the accuracy and automation level of river environmental management.

[0005] Given the shortcomings of the aforementioned solutions, there is an urgent need for an AI-based river solid waste monitoring and data management system to address these issues. Summary of the Invention

[0006] The purpose of this invention is to solve the technical problems mentioned in the background section and to provide an artificial intelligence-based solid waste monitoring and data management system.

[0007] The above-mentioned objective of the present invention is achieved through the following technical solution:

[0008] An AI-based solid waste monitoring and data management system includes:

[0009] The module includes an acquisition module, an encoder module, a feature enhancement module, an upsampling feature fusion module, a multi-branch detection head module, a dynamic weight fusion module, a classification and regression prediction module, and a result output module.

[0010] The acquisition module is used to acquire river surface image data of the river scene;

[0011] The encoder module is based on the RW-YOLOv11 architecture and includes a multi-scale backbone network composed of multiple C3K2_Sc feature extraction units, which is used to perform preliminary feature extraction and multi-scale feature generation on the input image.

[0012] The feature enhancement module utilizes the spatial normalization and channel dynamic selection mechanism in each C3K2_Sc unit to normalize the spatial differences of feature maps at different resolutions, and enhances the robustness of details in complex water surface scenes through residual connection.

[0013] The upsampling feature fusion module includes the SurfCAU feature enhancement unit, which is used to upsample and enhance low-resolution feature maps. It achieves the supplementary fusion of high-resolution features through local attention, cross-scale convolution, and dynamic weight calculation.

[0014] The multi-branch detection head module includes a SurfMSDFHead structure with four parallel branches, which are used to output feature maps of garbage targets at four different scales.

[0015] The dynamic weight fusion module calculates dynamic weights for cross-scale feature fusion based on the importance of features at each scale, and then weights and combines the features at each scale according to the weights to obtain high-resolution fused features.

[0016] Based on the above-mentioned fused features, the classification and regression prediction module outputs the classification probability of waste categories and the regression prediction of bounding box positions through a one-dimensional convolutional structure.

[0017] The results output module is used to output the monitoring results of solid waste in the river, including waste type, location information and confidence level.

[0018] As a preferred technical solution of the present invention, each C3K2_Sc module in the encoder module includes: a unit for performing channel grouping, group normalization and channel importance adjustment on the input image; wherein the normalization operation adjusts the scale of the features based on the mean difference and variance difference of the input features, and performs residual enhancement based on the learnable offset, so as to improve the feature representation ability of small target garbage under changes in river surface illumination and ripple interference.

[0019] As a preferred technical solution of the present invention, the feature enhancement module calculates a spatial weight map based on the feature map output by the C3K2_Sc module through a spatial reconstruction mechanism, which is used to suppress noise in the background region of the image and highlight the garbage target region; at the same time, the spatial weights are fused with the original feature map element by element using depth convolution and pixel-wise weighting, thereby obtaining a spatially refined and enhanced feature map.

[0020] As a preferred technical solution of the present invention, the SurfCAU module includes:

[0021] Convolutional units used to extract core information of low-resolution features;

[0022] An adaptive weight generation unit used to generate local dynamic convolutional kernel weights;

[0023] A deep convolutional unit used to apply the dynamic convolutional kernel to low-resolution features to reconstruct high-resolution local details;

[0024] The dynamic convolution kernel weights are normalized by Softmax and then used to weight different sampling positions, thereby achieving detail compensation and noise suppression for low-resolution features.

[0025] As a preferred technical solution of the present invention, the multi-branch detection head SurfMSDFHead includes four parallel output branches, each branch corresponding to a feature map at a different scale, for simultaneously outputting classification features and regression features of multi-scale river solid waste; wherein each branch includes three parts: a feature adjustment layer, a deep extraction layer and a prediction layer, to enhance the detection accuracy of small targets, dense targets and various types of river floating objects.

[0026] As a preferred technical solution of the present invention, the dynamic weight fusion module evaluates the importance of each branch feature based on the feature response of the four branches using a learnable weight generation unit, and performs weighted fusion of the branch features through normalized weights to obtain high-resolution fusion features of a uniform scale, thereby improving the robustness of the final classification and localization.

[0027] As a preferred technical solution of the present invention, the classification prediction module outputs the prediction probability for multiple waste categories based on fusion features through a convolutional structure. The category prediction probability is calculated based on Softmax and is used to distinguish multiple categories of river floating objects, including foam blocks, plastic bags, plastic bottles and other waste.

[0028] As a preferred technical solution of the present invention, the regression prediction module outputs the position parameters of the target bounding box based on the fusion features through the convolution structure, including the center coordinates and width and height information of the bounding box, and simultaneously outputs the target confidence score representing whether the bounding box is valid, which is used to determine the validity of the predicted box.

[0029] As a preferred technical solution of the present invention, the result output module adopts an IoU interval weighting strategy in the bounding box evaluation process, assigning different weights according to the IoU segment of the predicted box and the real box, in order to strengthen the penalty for low IoU samples and improve the model's localization stability for solid waste targets in river channels.

[0030] As a preferred technical solution of the present invention, the system is deployed on river inspection equipment, which collects real-time river surface images through unmanned boats or unmanned vessels, and runs the above-mentioned detection process on an embedded computing platform to realize real-time detection and positioning of various types of floating garbage on the river surface, which can support river cleaning, river surface monitoring and environmental management applications.

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] 1. This invention constructs an intelligent detection structure for river solid waste based on RW-YOLOv11. It introduces a C3K2_Sc feature extraction module into the encoder, effectively enhancing the edge and texture representation capabilities of floating debris through spatial attention and dynamic channel reconstruction mechanisms. This significantly improves the recognition of small target features against complex water surface backgrounds such as ripples, reflections, and strong light. In the multi-scale feature fusion section, this invention employs the SurfCAU water surface content-aware upsampling module, which dynamically generates appropriate upsampling weights based on different water surface textures. This achieves detail compensation for low-resolution features, reduces boundary blurring caused by traditional interpolation methods, and thus obtains higher-quality multi-scale features, providing richer semantic and structural information for subsequent detection.

[0033] 2. This invention employs a SurfMSDFHead multi-branch structure in the detection head network. By setting dynamic fusion weights for features at different scales, the higher-resolution branches contribute more to the detection of small targets, solving the detection difficulties caused by large differences in the size, uneven quantity, and dense accumulation of floating garbage. Simultaneously, this invention introduces a Focaler-IOU interval weighted loss function, dynamically adjusting the loss weights based on the overlap between the predicted and ground truth boxes. This enhances the training effect on low-overlap samples, effectively reducing localization errors in cases of garbage overlap, edge-to-edge, and partial occlusion, achieving more stable and accurate localization and classification of floating garbage in waterways.

[0034] 3. This invention achieves lightweight modeling while maintaining high-precision detection capabilities. The overall number of parameters and computational load are significantly lower than traditional detection models, making it suitable for embedded and onboard computing devices such as NVIDIA Jetson AGX Xavier, and meeting the operational requirements of real-time monitoring. Through data management and visualization modules, this invention can realize the associated storage of waste monitoring results, environmental information, and time-series data, and provide multi-dimensional analysis and governance suggestions. It can significantly improve the intelligence, automation, and data traceability capabilities of river solid waste monitoring, and has significant application value in scenarios such as water environment governance, smart water construction, and ecological monitoring. Attached Figure Description

[0035] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the 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.

[0036] Figure 1 is a diagram of the overall network structure of the present invention;

[0037] Figure 2 is a structural diagram of the C3K2_Sc module;

[0038] Figure 3 is the overall architecture diagram of the SurfCAU surface content awareness enhancement upsampling module;

[0039] Figure 4 is a structural diagram of the SurfMSDFHead module;

[0040] Figure 5 shows the results of a visual comparison experiment between the present invention and the state-of-the-art methods on a public dataset. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the embodiments and Figures 1-5. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0042] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. The described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0043] This implementation method unfolds with the "encoder-neck-decoder (detection head)" core link of the RW-YOLOv11 detection sub-architecture, clearly defining the composition and functional division of each module, enabling multi-scale feature extraction, fusion, and accurate prediction of floating debris in river channels, while connecting to the data management layer to achieve linkage between detection and data.

[0044] The encoder is used for the extraction and preliminary processing of features from river solid waste images, adapting to different types of river surface scenes, and achieving efficient capture of waste features through an original module. The input branch receives RGB images of river solid waste with a size of 640×640×3 collected by an unmanned surface vessel. The feature extraction backbone uses the C3K2_Sc module (as shown in Figure 2). The main branch of the water surface image contains 4 C3K2_Sc units, each of which undergoes dual dynamic optimization of spatial and channel dimensions to ensure the representation of multi-scale waste on the water surface. To ensure that the feature maps maintain appropriate resolution at different scales, a 3×3 convolution with a stride of 2 is set before each C3K2_Sc module (except for the first one) for downsampling, and finally outputs feature maps at 4 scales: 160×160×C, 80×80×C, 40×40×C, and 20×20×C, where C is the number of channels, initially 64, and the number of channels doubles with each downsampling.

[0045] The neck structure connects the encoder and the detection head, responsible for further fusion and detail enhancement of multi-scale features. Four SurfCAU water surface content-aware enhancement upsampling modules (as shown in Figure 3) are configured, corresponding one-to-one with the four scales output by the encoder. Starting from the minimum resolution of 20×20×C, SurfCAU scales up the feature map size by a factor of 2, fixing the upsampling factor s=2 and the kernel size K=3 to ensure proper reconstruction of feature details under complex river surface textures.

[0046] The detection head, as the final output unit, employs a four-branch SurfMSDFHead structure (as shown in Figure 4) and Focaler-IOU loss, enabling the model to classify and locate floating debris on the water surface at multiple scales. In each branch, a 1×1 convolution first unifies the channels to 256, followed by two 3×3 convolutions (with BN and ReLU) as deep extraction layers, and then into the prediction layers (classification and regression sub-layers). To enhance detection performance, a SurfMSDF dynamic weight unit is introduced before the four prediction layers, and four weights are obtained through Softmax. The sum of the four is 1, where For the high-resolution branch, the value is set to 0.5 to enhance the detection contribution of small garbage. The fusion formula is as follows:

[0047] .

[0048] in, The final feature map after fusion; : The feature map output by the i-th detection branch; Dynamic fusion weights, satisfying The classification sub-layer is through The convolution outputs the probabilities of 12 types of garbage (using Softmax activation), as shown in the following expression:

[0049] .

[0050] in, Let be the predicted probability of the c-th type of garbage at pixel (x, y); The classification convolutional layer predicts the value for class c; the denominator is the sum of the exponents of all 12 classes, implementing Softmax. The regression sub-layer outputs the bounding box coordinates through a 1x1 convolution. and confidence level The confidence score is calculated using the Sigmoid activation method.

[0051] ;

[0052] in, Target confidence level; This is the output of the regressive convolutional layer, used to represent the probability that the bounding box contains the target.

[0053] The loss is calculated using a weighted sum of Focaler-IOU and cross-entropy:

[0054] ;

[0055] in, For IoU-based interval localization loss; 0.7 represents the classification loss; 0.7 and 0.3 are empirical weighting coefficients.

[0056] The C3K2_Sc module is the core feature extraction unit of the RW-YOLOv11 encoder, enhancing feature representation through a dual dynamic approach of spatial optimization and channel filtering. The module first performs group normalization, normalizing the input feature map X (dimensions H x W x C) according to G=8 groups, calculated as follows:

[0057] .

[0058] in, Input feature map; , Let g be the mean and variance of the g-th group; , For learnable scale bias; For stable terms; G=8 is the number of groups. GN(X) passes through... Convolution (channels compressed to C / 4) and 3×3 convolution extract local features; the sum of the two is then passed through a Sigmoid function to obtain the spatial weight map. :

[0059] ;

[0060] in, Spatial attention map (HxW); The result of adjusting the channels to C / 4 for the 1×1 convolution output. This is the output of a 3×3 convolution; the Sigmoid function is used to normalize the weights.

[0061] Finally Element-wise multiplication with X yields the spatial refinement feature. The formula is:

[0062] .

[0063] Where x represents element-wise multiplication; These are features enhanced by spatial enhancement.

[0064] The dynamic channel reconstruction unit filters key feature channels and compresses redundant background channels through channel contribution analysis. Global average pooling is then performed to obtain the channel statistics s (dimension C), as shown in the formula:

[0065] ;

[0066] in, The value is the global average of the c-th channel; H and W are the height and width of the feature map, respectively. All coordinates are pixel coordinates.

[0067] The input is a lightweight MLP (two fully connected layers, with C / 16 hidden neurons), and the output is a dynamic compression ratio α (dimension C), as shown in the formula:

[0068] ;

[0069] in , The weights are for two fully connected layers; , For bias; , representing the compression weight of each channel. Based on α, the channels are divided into core channels (α≥0.5) and redundant channels (α<0.5). Both are processed by depthwise separable convolution, and then concatenated. The convolution is restored to the C channel and fused with the residual X:

[0070] ;

[0071] in, Core channel characteristics; Redundant channel features; Concat concatenates channel dimensions; dw3×3 is a depthwise separable 3×3 convolution; This is the final module output.

[0072] The SurfCAU module addresses the issue of blurred details in water surface content during upsampling. Through the combined action of the kernel prediction and feature reassembly stages, it reduces redundant computation while preserving the edge structure of floating debris in the river channel, thereby improving feature resolution and detail fidelity. SurfCAU first processes the input low-resolution feature map... (Dimensions H x W x C) are reduced to C / 4 through 1x1 convolution, resulting in:

[0073] ;

[0074] in, Low-resolution input features; The input feature map is dimensionality reduced by channel size; H, W, and C represent the height, width, and number of channels of the input feature map, respectively. Subsequently, SurfCAU generates a dynamic kernel weight matrix through 3×3 convolutions. (Dimension H×W×(K²×s²)), the formula is:

[0075] ;

[0076] Its dimensions are kernel size upsampling factor To ensure the interpretability and usability of kernel weights, softmax normalization is performed on the kernel weights at each position:

[0077] ;

[0078] in, The weight value is for the i-th kernel. : Represents the 36 weighting factors corresponding to each pixel; Softmax ensures that the weights are normalized so that the sum of all weights is 1.

[0079] The process then moves to the feature reassembly stage, where high-resolution features are reconstructed based on dynamic weights. For each pixel in F... Select Neighborhood features, expanded by s Blocks, generating local features (dimension) High-resolution pixels are generated through weighted aggregation. The formula is:

[0080] ;

[0081] The coordinate transformation is as follows:

[0082] ;

[0083] in, Low-resolution pixel coordinates; : The corresponding high-resolution coordinates; Offset, ranging from 0 to s-1; High-resolution features after reconstruction.

[0084] The reconstructed high-resolution features are compared with the previous scale by 1 The convolutionally adjusted features are added element-wise and fused to obtain the final output:

[0085] ;

[0086] in, : Feature map of the previous scale; The fused upsampled output. The SurfMSDFHead and Focaler-IOU together constitute the core structure of the detection head.

[0087] SurfMSDFHead enables dynamic fusion of multi-resolution features, while Focaler-IOU addresses the issue of inaccurate localization caused by overlapping small targets in river scenes. The input to SurfMSDFHead consists of four different resolution features from the neck network.

[0088] (160 ), (80 ), (40 ), (20 All were uniformly converted to 160 using bilinear interpolation. The formula is:

[0089] ,

[0090] ,

[0091] ,

[0092] .

[0093] Where Interpolate is a bilinear interpolation function; This is the result after adjusting features at different scales to the same resolution. The four resized features are concatenated and input into a fully connected layer, where they are then processed by Softmax to obtain the fusion weights.

[0094] ;

[0095] in, And satisfy:

[0096] .

[0097] Where W is the weight of the fully connected layer; b is the bias term; and β is the dynamic fusion weight vector, highlighting the importance of features at different scales.

[0098] The final fusion feature calculation is as follows:

[0099] ;

[0100] in, This is the final high-resolution feature map after fusion. The classification branch is... Convolution outputs the probability of 12 types of garbage:

[0101] ;

[0102] in, The probability of type c garbage To extract features using 3D convolution, is the output class logits; Softmax maps logits to probabilities.

[0103] The bounding box coordinates of the regression branch (1-4 channels) are obtained by sigmoid activation.

[0104] ;

[0105] in, The coordinates of the top left and bottom right corners of the prediction box; The probability of a garbage instance existing.

[0106] Focaler-IOU loss is used to evaluate the relationship between the predicted bounding box B and the ground truth bounding box B. IoU:

[0107] ;

[0108] Weight intervals are divided based on IoU:

[0109] ;

[0110] in, Weights are assigned to the IoU interval, causing the model to penalize more strongly in the low IoU range. The final Focaler-IOU loss formula is:

[0111] ;

[0112] Where y=2.5 represents the adjustment term, which is used to strengthen the penalty for low IoU samples.

[0113] After the overall design of the module was completed, in order to verify its effectiveness, an experimental dataset was constructed and experimental comparative analysis was conducted.

[0114] The experiment combined publicly available datasets with a self-made dataset to ensure the reliability and generalization of the experimental results. The publicly available dataset used was IWHR_AI_Label_Floater_V1, which contains 2000 images of solid waste in river channels, covering different lighting conditions such as sunny and cloudy days, as well as different terrain scenes such as plain rivers and mountain rivers. This dataset can be used to cross-validate the model's generalization ability under different publicly available scenarios.

[0115] The self-made dataset was collected from three typical rivers, totaling 7,100 images. It covers complex scenes such as garbage accumulation, water ripple interference, and underwater blurring. It labels 5 major categories and 12 subcategories of garbage, comprehensively covering the main forms and scenes of solid waste in rivers. The self-made dataset and the public dataset were mixed and divided into training set, validation set and test set in a ratio of 7:1:2. The training set of 5,000 images was used for model parameter learning, the validation set of 700 images was used for model hyperparameter tuning, and the test set of 1,400 images was used for evaluating the final performance of the model.

[0116] The experimental hardware environment included an NVIDIA RTX 4090 GPU, an Intel Core i9-13900K CPU, and 64GB of DDR5 memory for model training; the edge deployment device was an NVIDIA Jetson AGX Xavier with 32GB of RAM, used to test the real-time performance of the model on the edge device; the software environment was based on the Ubuntu 22.04 LTS operating system, and the deep learning framework used was PyTorch 2.1.0, with CUDA 12.1 acceleration enabled to ensure the efficiency of model training and inference.

[0117] Model training process: The model training parameters are set as follows: SGD is used as the optimizer, momentum is set to 0.9, and weight decay is set to 5e-4 to balance training stability and convergence speed; the initial learning rate is set to 0.01, and a cosine annealing learning rate scheduling strategy is adopted, with an annealing period of 300 and a minimum learning rate of 1e-5 to avoid overfitting in the later stages of training; the batch size is adjusted according to the hardware configuration, with a batch size of 16 for cloud training and a batch size of 4 for edge device fine-tuning; the training epochs are set to 300, with the first 100 epochs freezing the first 5 layers of the backbone network and training only the improved modules, and the last 200 epochs unfreezing the entire network for fine-tuning, saving the model weights every 20 epochs, and finally selecting the weight with the highest mAP on the validation set as the final model; at the same time, TensorRT 8.6 is used to perform INT8 quantization on the final model to further reduce the number of model parameters and computational load, and improve the deployment efficiency of edge devices.

[0118] Comparative Experiment Design: Several mainstream object detection models were selected as benchmarks, including Faster R-CNN, Dino, YOLOv5, YOLOv8, and YOLOv11. All models were trained and tested on the same dataset and with the same training parameters to ensure fairness in the comparison. Evaluation metrics included mAP@0.5, mAP@0.5:0.95, number of parameters, computational cost, and real-time detection frame rate to comprehensively measure the detection accuracy, lightweight nature, and real-time performance of the models.

[0119] Experimental Results Analysis: As shown in Tables 1 and 2, on the self-made test set, the mAP@0.5 of the model of this invention reached 0.857, an improvement of 11.5% compared to YOLOv11; the mAP@0.5:0.95 reached 0.759, an improvement of 19.8% compared to YOLOv11; and the small target AP reached 0.68, an improvement of 10 percentage points compared to YOLOv11. This indicates that the improved module effectively enhances the small garbage detection capability and improves the multi-scale garbage detection accuracy. On the public dataset IWHR_AI_Label_Floater_V1, the mAP@0.5 of the model of this invention reached 0.832, verifying the model's good generalization ability on different datasets.

[0120] The model of this invention has 8.9M parameters, a reduction of 4.2M compared to YOLOv11, maintaining its lightweight characteristics; the computational cost is 24.2G FLOPs, a reduction of 4.1G compared to YOLOv11. On the NVIDIA Jetson AGX Xavier edge device, the real-time detection frame rate reaches 48, an improvement of 11 compared to YOLOv11, meeting the real-time requirement of a detection frame rate of over 30 in unmanned surface vessel inspection scenarios.

[0121] Table 1. Quantitative experimental data of the model on a self-made dataset.

[0122] Methods (Models) mAP@0.5 mAP@0.5:0.95 Params(M) FLOPs(G) FPSFaster R-CNN 0.696 0.4772 0.735.420 Dino 0.726 0.5426 3.846.916 YOLOv5 0.703 0.5151 6.532.328 YOLOv8 0.715 0.5331 2.628.931 YOLOv11 0.742 0.5611 3.128.337 RW-YOLOv11 0.857 0.759 8.92 4.248 surface

[0123] Table 2. Quantitative experimental data of the model on the IWHR_AI_Label_Floater_V1 dataset.

[0124] Methods (Models) mAP@0.5 mAP@0.5:0.95 Params(M) FLOPs(G) FPSFaster R-CNN 0.651 0.4162 0.735.420 Dino 0.699 0.5256 3.846.916 YOLOv5 0.681 0.4971 6.532.328 YOLOv8 0.708 0.5201 2.628.931 YOLOv11 0.737 0.5411 3.128.337 RW-YOLOv11 0.832 0.741 8.924.248 surface

[0125] In complex scene tests, the model's mAP@0.5 was 0.813 under heavy rain, a decrease of only 4.4 percentage points compared to sunny weather. In the garbage overlap scene, the average localization error was 12.3 pixels, a reduction of 56.8% compared to YOLOv11. This indicates that the model of this invention has strong robustness in complex river environments and can effectively cope with harsh scenarios such as heavy rain, underwater blur, and garbage overlap, and stably complete the garbage detection task.

[0126] This invention designs a high-performance end-to-end detection algorithm, RW-YOLOv11, for river solid waste and implements full-process data management, enabling high-precision, high-efficiency detection and intelligent management of river solid waste. It provides a practical technical solution for water environment governance and has significant engineering application value and promotion prospects.

[0127] To visually compare the detection performance of the models, a visualization experiment was conducted between RW-YOLOv11 and YOLOv11, and the results are shown in Figure 5. In a simple scene (first row), RW-YOLOv11 not only accurately selected the targets for "foam board" (90% confidence) and "plastic bottle" (95% confidence), but also achieved significantly higher confidence levels than YOLOv11 for "foam board" (86% confidence) and "plastic bottle" (79% confidence). In a complex scene with multiple targets (second row), RW-YOLOv11 could simultaneously identify multiple different types of waste, and the confidence levels for each target were high. In contrast, YOLOv11 had lower confidence levels for some targets, and even exhibited missed detections and misclassifications. This indicates that the RW-YOLOv11 proposed in this invention outperforms YOLOv11 in terms of recognition accuracy and visual performance for small and multiple targets in actual floating waste detection on water surfaces, and is better suited to meet the visualization needs of river solid waste inspection.

[0128] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A solid waste monitoring and data management system based on artificial intelligence, characterized in that, include: The system comprises an acquisition module, an encoder module, a feature enhancement module, an upsampling feature fusion module, a multi-branch detection head module, a dynamic weight fusion module, a classification and regression prediction module, and a result output module. The acquisition module acquires river surface image data for river scenes. The encoder module, based on the RW-YOLOv11 architecture, includes a multi-scale backbone network composed of multiple C3K2_Sc feature extraction units, used for preliminary feature extraction and multi-scale feature generation from the input image. The feature enhancement module utilizes spatial normalization and dynamic channel selection mechanisms in each C3K2_Sc unit to normalize the spatial differences of feature maps at different resolutions and enhances the robustness to details in complex water surface scenes through residual connections. The upsampling feature fusion module includes a SurfCAU feature enhancement unit. The system employs a multi-branch detection head module to upsample and enhance low-resolution feature maps, supplementing and fusing high-resolution features through local attention, cross-scale convolution, and dynamic weight calculation. The multi-branch detection head module includes a SurfMSDFHead structure with four parallel branches, each outputting feature maps of waste targets at four different scales. The dynamic weight fusion module calculates dynamic weights for cross-scale feature fusion based on the importance of features at each scale, and weights the features at each scale to obtain high-resolution fused features. The classification and regression prediction module, based on the fused features, outputs the classification probability of waste categories and the regression prediction of bounding box positions through a one-dimensional convolution structure. The result output module outputs the monitoring results of solid waste in the river channel, including waste category, location information, and confidence level.

2. The artificial intelligence-based solid waste monitoring and data management system according to claim 1, characterized in that, Each C3K2_Sc module in the encoder module includes: a unit for performing channel grouping, group normalization, and channel importance adjustment on the input image; wherein the normalization operation adjusts the scale of the features based on the mean difference and variance difference of the input features, and performs residual enhancement based on the learnable offset to improve the feature representation capability of small target garbage under changes in river surface illumination and ripple interference.

3. The artificial intelligence-based solid waste monitoring and data management system according to claim 1, characterized in that, The feature enhancement module calculates a spatial weight map based on the feature map output by the C3K2_Sc module through a spatial reconstruction mechanism. This map is used to suppress noise in the background region of the image and highlight the garbage target region. At the same time, the spatial weights are fused with the original feature map element by element using depthwise convolution and pixel-wise weighting to obtain a spatially refined and enhanced feature map.

4. The artificial intelligence-based solid waste monitoring and data management system according to claim 1, characterized in that, The SurfCAU module includes: a convolutional unit for extracting core information of low-resolution features; an adaptive weight generation unit for generating local dynamic convolutional kernel weights; and a depth convolutional unit for applying the dynamic convolutional kernel to low-resolution features to reconstruct high-resolution local details. The dynamic convolutional kernel weights are normalized by Softmax and used to weight different sampling positions, thereby achieving detail compensation and noise suppression for low-resolution features.

5. The artificial intelligence-based solid waste monitoring and data management system according to claim 1, characterized in that, The multi-branch detection head SurfMSDFHead includes four parallel output branches, each corresponding to a feature map at a different scale, which are used to simultaneously output the classification and regression features of multi-scale river solid waste. Each branch consists of three parts: a feature adjustment layer, a deep extraction layer, and a prediction layer, to enhance the detection accuracy of small targets, dense targets, and various types of floating objects in the river.

6. The artificial intelligence-based solid waste monitoring and data management system according to claim 1, characterized in that, The dynamic weight fusion module evaluates the importance of each branch feature based on the feature response of the four branches using a learnable weight generation unit, and performs weighted fusion of the branch features through normalized weights to obtain high-resolution fusion features of a uniform scale, thereby improving the robustness of the final classification and localization.

7. The artificial intelligence-based solid waste monitoring and data management system according to claim 1, characterized in that, The classification prediction module is based on fused features and outputs the prediction probability for multiple waste categories through a convolutional structure. The category prediction probability is calculated based on Softmax and is used to distinguish multiple categories of floating debris in the river, including foam blocks, plastic bags, plastic bottles and other waste.

8. The artificial intelligence-based solid waste monitoring and data management system according to claim 1, characterized in that, The regression prediction module, based on fused features, outputs the positional parameters of the target bounding box through a convolutional structure, including the center coordinates and width and height information of the bounding box, and simultaneously outputs the target confidence score representing whether the bounding box is valid, which is used to determine the validity of the predicted box.

9. The artificial intelligence-based solid waste monitoring and data management system according to claim 1, characterized in that, The result output module employs an IoU interval weighting strategy during the bounding box evaluation process. Different weights are assigned based on the IoU interval between the predicted box and the ground truth box to strengthen the penalty for low IoU samples and improve the model's localization stability for solid waste targets in river channels.

10. The artificial intelligence-based solid waste monitoring and data management system according to claim 1, characterized in that, The system is deployed on river inspection equipment, collects real-time river surface images through unmanned boats or unmanned vessels, and runs the above detection process on an embedded computing platform to achieve real-time detection and location of various types of floating garbage on the river surface, supporting applications in river cleaning, river surface monitoring and environmental management.

Images