Multi-scene micro garbage real-time detection method and system

Abstract

The invention discloses a multi-scene micro garbage real-time detection method and a multi-scene micro garbage real-time detection system, and aims to solve the defects of an existing garbage detection method in the aspects of small target identification, environmental adaptability and system integration. According to the method, on the basis of an improved YOLOv8 architecture, a Transform module and an efficient channel attention mechanism are fused, and the tiny target feature extraction and multi-scale information fusion capability is enhanced. The system adopts dynamic resolution reasoning, INT8 quantitative compression and time-space domain post-processing, so that the detection precision and the reasoning efficiency are remarkably improved. Experiments prove that the system effectively improves the small target recall rate and environmental adaptability in a garbage detection task, comprises eight corresponding modules, realizes full-process automatic detection, has relatively high real-time performance and automation level, is suitable for various application scenes such as smart city environment monitoring, and has important practical application value.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of computer vision and intelligent perception technology, and in particular to a multi-scene micro garbage real-time detection method and system. BACKGROUND

[0002] In recent years, the rapid development of computer vision and artificial intelligence technology has promoted the wide application of target detection in the fields of security monitoring, autonomous driving, industrial detection and smart city. However, in the aspect of multi-scene micro garbage real-time detection, the existing technology still faces many challenges and needs to be broken through.

[0003] Micro garbage target detection, as a key link in the environmental management of smart city, is extremely easy to be missed or misjudged in complex environments such as urban roads, squares and parks due to its small target size and weak feature expression. Traditional detection methods often have difficulty in considering garbage targets of different sizes, shapes and materials, especially in actual conditions such as light changes, shielding interference and complex backgrounds, and the detection robustness and accuracy are significantly affected.

[0004] Although the current mainstream detection models such as feature pyramid network (FPN) and path aggregation network (PANet) have improved multi-scale perception ability, the detection performance is still not ideal in the scene where large garbage cans and micro cigarette butts with extreme scale changes exist at the same time. In view of the above problems, some researches introduce innovative methods such as dynamic attention mechanism and adaptive feature enhancement to optimize the detection effect of small targets. However, these methods generally have defects such as high computational complexity, strong hardware dependence and insufficient real-time performance, which are difficult to meet the efficient and real-time detection needs of micro garbage targets in multi-scene.

[0005] Therefore, it is urgent to develop a new micro garbage detection method and system for urban multi-scene with high precision and high real-time performance, which can break through the bottleneck of existing technology and realize accurate identification and intelligent monitoring of micro garbage targets in complex environments. This has important significance for improving the efficiency of urban environmental governance and promoting the construction of smart city, and belongs to the frontiers of intelligent environmental monitoring and target identification technology. SUMMARY

[0006] Therefore, the technical problem to be solved by the present application is to provide a multi-scene micro garbage real-time detection method and system, which aims to solve the deficiencies of existing garbage detection methods in small target identification, environmental adaptability and system integration.

[0007] To achieve the above purpose, the technical scheme adopted by the present application is as follows: A multi-scene micro garbage real-time detection method, comprising the following steps: Step S1, data collection and environmental self-adaptive preprocessing. Environmental image data covering urban roads, squares, parks and other multi-source scenes are collected, focusing on covering micro-garbage targets. The original data is subjected to noise suppression, brightness normalization, spatial geometry correction and multi-scale random enhancement, and sample balanced collection is performed according to the micro-target distribution characteristics, so as to improve the data quality and sample diversity.

[0008] Step S2, innovative hybrid feature extraction network design. A backbone network is constructed by fusing a windowed visual Transformer module and an efficient channel attention (ECA) mechanism, realizing cross-space long-distance context modeling and important feature adaptive enhancement. A variable kernel convolution (AKConv) is embedded at each feature layer to dynamically adjust the receptive field to adapt to targets of different scales, significantly improving the sensitivity to micro-garbage targets and the adaptability to complex environments.

[0009] Step S3, dynamic multi-scale feature fusion and information flow. A bidirectional feature pyramid fusion module (BiFPN) is designed, integrating spatial dimension integration (SDI), adaptive feature fusion (ASF) and cross-scale lightweight feature interaction (CCFM) mechanisms, realizing collaborative compensation and fine-grained information fusion of multi-level features, and strengthening micro-target feature expression and global semantic complementation.

[0010] Step S4, attention-guided feature screening and regulation. Spatial attention weight maps are generated through parallel pooling operations, and the dynamic gradient modulation algorithm (DGM) is used to regulate the micro-target feature propagation path in real time, effectively alleviating the small target feature disappearance and weak expression problem. A lightweight channel-spatial attention module (CBAM) is introduced to enhance the model's response ability to micro-target regions in complex environments and overall robustness.

[0011] Step S5, decoupled multi-branch detection head and optimization. A decoupled multi-branch detection head is developed, with an innovative dedicated high-resolution small target detection branch, combined with a scale-adaptive sample allocation algorithm to dynamically adjust the positive and negative sample ratio, significantly improving small target recall rate and positioning accuracy. SIoU loss function is used to optimize the bounding box regression, taking into account the directionality and shape matching, and improving the positioning performance.

[0012] Step S6, inference optimization and low-power deployment. Dynamic resolution inference mechanism and INT8 quantization compression technology are integrated, and the input resolution is automatically switched according to the scene target density, realizing efficient and low-power inference on edge devices, with model size reduction and inference speed improvement better than existing technologies.

[0013] In addition, the present application also provides a multi-scene micro-garbage real-time detection system, comprising: The data acquisition and preprocessing module integrates a high-resolution industrial camera and a multi-spectral light supplement device to obtain micro-litter image data in multiple environments such as urban roads, squares, parks, and beaches. The preprocessing process includes environmental adaptive noise suppression, brightness normalization, spatial geometry correction, and multi-scale random enhancement. Combined with a sample balanced collection strategy, the data quality and diversity of micro targets are significantly improved.

[0014] The innovative feature extraction network module integrates windowed visual Transformer units and efficient channel attention (ECA) mechanisms in the backbone to realize cross-space long-distance context modeling and adaptive enhancement of key features. The variable kernel convolution (AKConv) embedded in each feature layer dynamically adjusts the receptive field to adapt to different scale litter targets, effectively improving the sensitivity to micro-litter and the adaptability to complex environments.

[0015] The multi-scale feature fusion module uses a bidirectional feature pyramid (BiFPN) architecture, innovatively integrating spatial dimension integration (SDI), adaptive feature fusion (ASF), and cross-scale lightweight feature interaction (CCFM) mechanisms to realize collaborative compensation and fine-grained information flow of multi-level features, strengthening the semantic expression of micro targets and global complementary ability.

[0016] The attention-guided feature selection module generates a spatial attention weight map through parallel operations of average pooling and maximum pooling, and uses a dynamic gradient modulation algorithm (DGM) to real-time regulate the micro-litter feature propagation path, alleviating the small target feature disappearance and weak expression problem. A lightweight channel-spatial attention module (CBAM) is introduced to enhance the system's response ability to micro-litter areas in complex environments and overall robustness.

[0017] The decoupled multi-branch detection head module innovatively sets a dedicated high-resolution micro-litter detection branch, dynamically adjusts the positive and negative sample ratio through a scale-adaptive sample allocation algorithm, significantly improves the micro-litter recall rate and positioning accuracy. The SIoU loss function is used to optimize the bounding box regression, considering the directionality and shape matching to improve the positioning performance.

[0018] The inference optimization and low-power deployment module integrates a dynamic resolution inference mechanism and INT8 quantization compression technology, automatically switches the input resolution according to the scene target density, realizes efficient and low-power inference on edge devices, and reduces the model size and improves the inference speed, which is better than existing technologies.

[0019] The post-processing and false alarm filtering module uses trajectory continuity verification and multi-modal fusion strategies, combines temporal information filtering and depth / thermal imaging data fusion to effectively reduce false alarm rate and improve detection reliability in complex environments.

[0020] The visual application interface module supports a RESTful API, a real-time visualization interface, and automatic alarm and statistical report generation functions, meets multi-platform integration and smart city management requirements, and realizes automatic detection, real-time monitoring, and efficient management of micro garbage targets in a city environment.

[0021] As a preferred scheme of the present application, the mixed feature extraction network integrates a windowed visual Transformer module and an efficient channel attention (ECA) mechanism in the backbone structure, and embeds an adaptive kernel convolution (AKConv) at each level of the feature layer, realizes cross-spatial long-distance context modeling and multi-scale feature adaptive enhancement, and effectively improves the sensitivity and representation ability of the model to micro garbage targets in complex environments.

[0022] As a preferred scheme of the present application, the multi-scale feature fusion module adopts a bidirectional feature pyramid (BiFPN) architecture, innovatively integrates spatial dimension integration (SDI), adaptive feature fusion (ASF), and cross-scale lightweight feature interaction (CCFM) mechanisms, and through dynamic compensation and fine-grained information flow of multi-level features, strengthens the semantic expression and global complementary ability of micro targets.

[0023] As a preferred scheme of the present application, the attention-guided feature screening mechanism generates a spatial attention weight map through parallel operation of average pooling and maximum pooling, and combines a dynamic gradient modulation algorithm (DGM) to real-time regulate the micro target feature propagation path, and simultaneously introduces a lightweight channel-spatial attention module (CBAM), which significantly enhances the robustness of the model in complex environments and the response ability of the micro target region.

[0024] Compared with the prior art, the present application has the following technical effects: 1) Significantly improve the micro target detection accuracy. By introducing a windowed visual Transformer module, an efficient channel attention (ECA) mechanism, and an adaptive kernel convolution (AKConv) in the network structure, deep modeling of cross-spatial long-distance context information and adaptive enhancement of multi-scale features are realized, which effectively enhances the sensitivity and representation ability of the model to micro garbage targets, and significantly reduces the miss detection and false detection rates.

[0025] 2) Enhance system robustness and generalization performance. Innovatively adopt a bidirectional feature pyramid (BiFPN) fusion architecture, integrate spatial dimension integration (SDI), adaptive feature fusion (ASF), and cross-scale lightweight feature interaction (CCFM), realize dynamic compensation and fine-grained information flow of multi-level features, so that the system can effectively cope with detection challenges in extreme scale changes, uneven target distribution, and complex urban environments.

[0026] 3) Improve the response ability of small target area. An attention-guided feature selection and regulation mechanism is proposed, which generates a spatial attention weight map through parallel average pooling and maximum pooling, combines with a dynamic gradient adjustment algorithm (DGM) to optimize the small target feature propagation path in real time, and introduces a lightweight channel and spatial attention module (CBAM) to further improve the overall robustness and response ability of the model in complex scenes.

[0027] 4) Improve positioning accuracy and recall rate. The detection head adopts a decoupled multi-branch architecture, adds a dedicated high-resolution small target detection branch, and dynamically adjusts the positive and negative sample ratio with a scale adaptive sample allocation algorithm to significantly improve the small target recall rate and positioning accuracy. The SIoU loss function is used to optimize the boundary box regression, which takes into account the directionality and shape matching to ensure the high accuracy and stability of the detection results.

[0028] 5) Optimize inference efficiency and system integration. The system integrates dynamic resolution inference mechanism and INT8 quantization compression technology, intelligently switches the input resolution according to the scene target density, realizes efficient and low-power inference on edge devices, significantly reduces the model size, and significantly improves the inference speed, fully meeting the practical application requirements of garbage automatic identification and real-time monitoring in smart city environment.

[0029] Of course, any product implementing the present application does not necessarily need to achieve all the technical effects described above. BRIEF DESCRIPTION OF DRAWINGS

[0030] The accompanying drawings described herein are used to provide a further understanding of the present application, and form a part of the present application. The illustrative embodiments of the present application and their descriptions serve to explain the present application, and do not constitute an improper limitation on the present application. In the drawings: Figure 1 is the overall system architecture flowchart of the present application; Figure 2 is the mixed feature extraction network structure of the present application; Figure 3 is the multi-scale feature fusion module structure of the present application; Figure 4 is the intelligent detection system of the multi-scene micro garbage real-time detection method of the present application; Figure 5 is the application schematic diagram of the multi-scene micro garbage real-time detection system of the present application; DETAILED DESCRIPTION

[0031] The embodiments of the present application will be described in detail below with the aid of the drawings and examples, so that the realization process of how the present application applies technical means to solve technical problems and achieve technical effects can be fully understood and implemented.

[0032] The system hardware configuration is that: it can be deployed in a high-performance computing platform, an industrial high-resolution camera is selected for the image acquisition device, a suitable light source is equipped, and the storage device includes a high-speed SSD and a large-capacity HDD.

[0033] The software environment of the application is that: the system supports Linux operating system, adopts mainstream deep learning framework PyTorch, and the programming language is Python.

[0034] The application provides a multi-scene micro garbage real-time detection method, as shown in Figure 1 The method mainly comprises the following steps: Step S1: multi-source environment image data acquisition and preprocessing In the first link of the garbage detection system, the system deploys multi-source camera equipment (such as high-resolution RGB camera and thermal imager) to collect data in all time periods and multiple angles in complex scenes such as urban roads, squares and parks, covering various distribution states of micro garbage targets. In order to improve the data quality, the preprocessing process includes environment adaptive noise suppression, brightness normalization and spatial geometric correction, which effectively eliminates the interference signals caused by light changes, equipment differences and scene complexity. In the aspect of data enhancement, the system adopts a multi-scale random transformation strategy, sets the scaling factor The value range is , and the calculation formula of the enhanced image is

[0035] Among them, Indicates the enhanced image, Indicates the original image, Indicates that the size is scaled by Through random adjustment of brightness, contrast and saturation, the robustness of the model under color change conditions is further enhanced. Finally, all collected data are reasonably divided into training set, verification set and test set to ensure that the target scale distribution in each set is balanced, laying a solid data foundation for subsequent feature extraction and model training.

[0036] Step S2: innovative hybrid feature extraction network design In the feature extraction stage, the system adopts an innovative hybrid backbone network, which deeply integrates the visual Transformer module and the convolutional neural network (CNN) structure, realizes efficient modeling of long-distance context information across space, and as shown in Figure 2Swin Transformer Block is embedded after the C2f module to capture multi-scale semantic associations through a windowed attention mechanism, enhancing the model's understanding of complex scenes. Meanwhile, an efficient channel attention (ECA) module is integrated into each Bottleneck structure to achieve adaptive enhancement of important features using global average pooling and one-dimensional convolution operations, with the formula

[0037] wherein, represents an activation function, represents global average pooling of input features , and is a one-dimensional convolution operation. The backbone network embeds variable kernel convolution (AKConv) at each feature layer to dynamically adjust the receptive field of the convolution kernel according to the scale of the input features, effectively capturing fine-grained information of different size targets. The above structures work together to greatly enhance the model's sensitivity to small garbage targets and adaptability to complex environments.

[0038] Step S3: Design of dynamic multi-scale feature fusion module In the multi-scale feature fusion stage, the system uses a bidirectional feature pyramid (BiFPN) structure combined with spatial dimension integration (SDI), adaptive feature fusion (ASF), and cross-scale feature interaction (CCFM) mechanisms to achieve dynamic compensation and information flow of multi-level features, as shown in Figure 3 BiFPN fuses features of different scales through weighted fusion, using an adaptive weight allocation formula

[0039] wherein, represents the layer fused feature map, represents the input feature, and the adaptive fusion weight. The SDI module focuses on local detail areas in the image through spatial dimension reduction operations, improving the recognition of small targets; the ASF mechanism smooths the information flow between features of different scales through adaptive fusion strategies, alleviating the feature discontinuity caused by scale changes. The lightweight CCFM design promotes effective interaction between high and low level features, further strengthening the propagation and expression of small target features. The overall fusion module not only significantly improves the model's adaptability to extreme scale change scenes, but also provides a more rich and reliable multi-scale feature basis for the subsequent detection head.

[0040] Step S4: Attention-guided feature selection mechanism In the feature selection and tuning stage, the system proposes an attention-guided feature selection mechanism. This mechanism generates a spatial attention weight map through parallel pooling operations, effectively improving the model's response to small target regions. Specifically, the average pooling and max pooling results are concatenated, and then convolutional and sigmoid activation are applied to obtain the spatial attention weights. The calculation formula is as follows:

[0041] in, Represents spatial attention weights. For the Sigmoid activation function, For convolution operations, Indicates channel splicing. and respectively for feature maps Global average pooling and max pooling are employed. Combined with a dynamic gradient adjustment mechanism (DGM), the system optimizes the flow of small target features in real time within the feature pyramid network. The feature adjustment formula is as follows:

[0042] in, The characteristics after regulation, and These represent low-level and high-level features, respectively. The system also introduces a lightweight channel spatial attention module (CBAM) to jointly model attention weights in both channel and spatial dimensions, effectively mitigating the problems of feature vanishing and weak representation of small targets, and significantly enhancing detection stability in complex environments. Step S5: Design and Optimization of Decoupled Multi-Branch Detection Head During the detection head design and optimization phase, the system adopts a decoupled multi-branch detection head architecture, separating classification and regression tasks to reduce mutual interference between tasks. For small target detection, the system adds a dedicated high-resolution feature branch, effectively improving the ability to recognize small-scale targets. The detection head output includes class probabilities and bounding box coordinates, calculated using the following formulas:

[0043] in, Represents class probability, For normalized activation functions, and These represent the weights and biases of the classification branches, respectively. As input features, Represents the bounding box coordinates. and These represent the weights and biases of the regression branch, respectively. The loss function employs a multi-task weighted design, with the overall optimization objective as follows:

[0044] in, For classifying losses, To regress the loss, For IoU loss, These represent the weight coefficients of each loss branch. Through decoupling structure and multi-branch optimization, the system's detection accuracy and robustness for small debris targets are effectively improved.

[0045] Step S6: Model Inference and Deployment Optimization In the model inference and deployment phases, the system efficiently optimizes the model structure and computation process based on actual application scenarios. During inference, the Non-Maximum Suppression (NMS) algorithm is used to filter redundant candidate boxes, and the final detection result is obtained through the following formula:

[0046] in, This represents the final output target bounding box. This indicates a nonmaximum suppression operation. For the coordinates of all candidate boxes, For the corresponding category probability, This represents the IoU threshold. To meet the deployment requirements of edge devices, the system employs model quantization and pruning techniques to reduce the number of model parameters and computational complexity. Quantization maps floating-point weights to low-order integers, specifically as follows:

[0047] in, The quantified weights, For the original weights, To quantize the step size, the above optimizations enable the system to achieve efficient and low-latency target detection inference, ensuring the model's real-time performance and stability in real-world environments.

[0048] This invention also provides an intelligent detection system based on a real-time detection method for micro-waste in multiple scenarios, such as... Figure 4 As shown, it includes: Module 1: Data Acquisition and Environmental Adaptive Preprocessing Module. Its components include an industrial-grade high-resolution camera (resolution ≥ 5 megapixels) and a multispectral illumination device. Specifically, the system acquires environmental image data covering tiny litter targets in multi-source scenarios such as urban roads, squares, parks, and beaches through automatic acquisition and synchronous triggering mechanisms. The preprocessing workflow integrates environmental adaptive noise suppression, brightness normalization, spatial geometric correction, and multi-scale random enhancement operations, combined with a sample equalization acquisition strategy to improve data quality and target distribution diversity, laying the foundation for subsequent feature extraction.

[0049] Module 2: An innovative feature extraction network module, whose components include a backbone network structure (embedded with a windowed visual Transformer module and a Swing Transformer Block), an efficient channel attention (ECA) mechanism, and variable kernel convolution (AKConv) units. The backbone network integrates a Swing Transformer Block after the C2f module, achieving multi-scale semantic association modeling through the windowed attention mechanism. ECA and AKConv modules are embedded within each Bottleneck structure to dynamically adjust the receptive field, enhancing the model's sensitivity to small debris targets and its adaptability to complex environments.

[0050] Module 3: Multi-scale feature fusion module, adopting a bidirectional feature pyramid (BiFPN) architecture, integrates spatial dimensionality reduction ensemble (SDI), adaptive feature fusion (ASF), and cross-scale lightweight feature interaction (CCFM) mechanisms. Through dynamic compensation of multi-level features and fine-grained information flow, it achieves semantic representation of small targets and global complementarity, significantly improving the system's detection robustness in extreme scale variation scenarios.

[0051] Module 4: Attention-guided feature selection module, including a spatial pooling unit, a dynamic gradient adjustment algorithm (DGM), and a lightweight channel spatial attention module (CBAM). It generates a spatial attention weight map through parallel operations of average pooling and max pooling, and combines DGM to optimize the propagation path of small target features in real time, effectively mitigating the problems of feature vanishing and weak representation of small targets. The CBAM module, through joint channel and spatial dimension modeling, further enhances the responsiveness of micro-garbage regions in complex environments and the overall robustness of the system.

[0052] Module 5: Decoupled multi-branch detection head module, whose structure includes a dedicated high-resolution micro-garbage detection branch, a scale-adaptive sample allocation algorithm unit, and an SIoU loss function regression optimization unit. Through decoupling design, classification and regression tasks are separated, improving the recall rate and localization accuracy of small targets. A multi-task weighted loss function is employed, balancing class recognition and bounding box regression to ensure the accuracy and stability of the detection results.

[0053] Module 6: Inference Optimization and Low-Power Deployment Module, integrating a dynamic resolution inference mechanism and INT8 quantization compression technology. It intelligently switches the input resolution based on the target density of the scene, enabling efficient and low-power inference on edge devices, reducing the number of model parameters and computational complexity, and ensuring the real-time performance and stability of the system in practical applications.

[0054] Module 7: Post-processing and false alarm filtering module, including a trajectory continuity verification unit, multi-modal fusion strategy (temporal information filtering, depth / thermal imaging data fusion), and false alarm screening mechanism. This module effectively reduces the false alarm rate in complex environments, improves detection reliability, and provides data support for subsequent intelligent management and alarms.

[0055] Module 8: Visual Application Interface Module, providing a RESTful API, real-time visualization interface, and automatic alarm and statistical report generation functions, supporting multi-platform integration and smart city management needs. The system can output detection results, target location, category information, and confidence levels in JSON format or visual charts, meeting the needs of various application scenarios such as environmental monitoring, intelligent alarms, and data analysis.

[0056] The various modules of this system work collaboratively to achieve comprehensive automatic detection, real-time monitoring, and efficient management of micro-sized waste targets in urban environments. It boasts technological advantages such as high-precision small target recognition, low false alarm rate, and high inference speed, making it suitable for various application scenarios including smart city environmental governance, public safety, and intelligent monitoring.

[0057] This invention's system can be applied to the detection of various types of waste in various scenarios, such as packaging bags, cigarette butts, aluminum cans, bottle caps, etc., in different scenarios. Figure 5 As shown, the workflow is as follows: First, the system uses a high-resolution RGB camera and thermal imaging equipment to collect data on litter targets from multiple sources, including urban roads, squares, and parks, around the clock and from multiple angles, focusing on micro-litter items such as cigarette butts, bottle caps, and wrappers. The collected image data is automatically fed into a preprocessing module, where noise suppression, brightness normalization, and spatial geometric correction are performed sequentially. A multi-scale random enhancement strategy is employed to optimize sample distribution and improve the quality and balance of small target data. The preprocessed images then enter the feature extraction stage. The backbone network integrates a windowed visual Transformer module with an efficient channel attention (ECA) mechanism to enhance the long-range contextual modeling capability for micro-litter targets. Subsequently, the feature fusion module adopts a bidirectional feature pyramid (BiFPN) structure, integrating spatial dimensionality reduction ensemble (SDI), adaptive feature fusion (ASF), and cross-scale lightweight feature interaction (CCFM) to achieve dynamic compensation of multi-level features and fine-grained information flow, significantly enhancing the semantic expression of micro-targets. In the detection head stage, the system adopts a decoupled multi-branch architecture, adding a dedicated high-resolution small target detection branch, and combining it with a scale-adaptive sample allocation algorithm to dynamically increase the proportion of positive samples for small targets. Bounding box regression optimization employs the SIoU loss function, balancing directionality and shape matching to further improve positioning accuracy and detection stability. The inference module integrates a dynamic resolution mechanism and INT8 quantization compression technology, automatically switching input resolution based on scene target density to achieve efficient and low-power inference on edge devices. The post-processing stage utilizes trajectory continuity verification and multimodal fusion strategies, combining temporal information filtering and thermal imaging data fusion to significantly reduce false alarm rates and improve detection reliability in complex environments. Finally, the system's visualization panel displays the detection results intuitively, including detailed information such as waste target type, location, and confidence level, and supports automatic alarm and data statistics functions, providing strong technical support for urban environmental governance and intelligent monitoring. Practical applications demonstrate that the system can efficiently identify various micro-sized waste targets, significantly improving detection accuracy and environmental adaptability, fully validating the innovation and practical value of this invention.

[0058] The foregoing description illustrates and describes several preferred embodiments of the present invention. However, as previously stated, it should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the inventive concept described herein through the foregoing teachings or techniques or knowledge in related fields. Any modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A method for real-time detection of micro-waste in multiple scenarios, characterized in that, The steps include the following: S1. Collect and preprocess multi-source environmental image data, which covers small litter targets in complex scenes such as city streets, parks, and beaches. The preprocessing process includes environmental adaptive noise suppression, brightness normalization, spatial geometric correction and multi-scale random enhancement, and sample equalization is performed to target the distribution characteristics of small targets. S2. Construct an innovative hybrid feature extraction network. The network integrates a windowed visual Transformer module and an efficient channel attention (ECA) mechanism in the backbone to achieve cross-space long-distance context modeling and adaptive enhancement of important features. Furthermore, it embeds a variable kernel convolution (AKConv) in each feature layer to dynamically adjust the receptive field to adapt to targets of different scales. S3. Design a bidirectional feature pyramid fusion module with adaptive weight allocation, innovatively integrating spatial dimensionality reduction fusion (SDI), adaptive feature fusion (ASF), and cross-scale lightweight feature interaction (CCFM) to achieve collaborative compensation of multi-level features, information flow, and fine-grained small target feature enhancement. S4. An attention-guided feature selection mechanism is proposed. An attention graph is generated through parallel pooling operations and combined with a dynamic gradient adjustment algorithm (DGM) to adjust the propagation path of small target features in real time, effectively alleviating the problems of small target feature disappearance and weak expression. At the same time, a lightweight channel space attention module (CBAM) is introduced to improve the overall robustness. S5. Develop a decoupled multi-branch detection head, innovatively set up a dedicated high-resolution small target detection branch, and combine it with a scale-adaptive sample allocation algorithm to dynamically adjust the ratio of positive and negative samples, which significantly improves the recall rate and positioning accuracy of small targets. The SIoU loss function is used to optimize bounding box regression, taking into account both directionality and shape matching. S6 integrates a dynamic resolution inference mechanism and INT8 quantization compression technology, automatically switching the input resolution according to the target density of the scene, achieving efficient and low-power inference on edge devices, and the model size reduction and inference speed improvement are both superior to existing technologies.

2. The method according to claim 1, characterized in that, The hybrid feature extraction network introduces a Swing Transformer Block after the C2f module and integrates efficient channel attention (ECA) and variable kernel convolution (AKConv) within each Bottleneck structure to improve the model's sensitivity to small targets and environmental adaptability.

3. The method according to claim 1, characterized in that, The multi-scale feature fusion module adopts a bidirectional feature pyramid (BiFPN) structure and combines spatial dimensionality reduction integration (SDI), adaptive feature fusion (ASF), and cross-scale feature interaction (CCFM) mechanisms to achieve dynamic compensation and fine-grained information fusion of multi-level features.

4. The method according to claim 1, characterized in that, The attention-guided feature selection mechanism generates a spatial attention weight map through parallel operations of average pooling and max pooling, and injects a dynamic gradient adjustment mechanism (DGM) into the feature pyramid network. Combined with a lightweight channel spatial attention module (CBAM), it improves the model's responsiveness to small target regions and its robustness in complex environments.

5. The method according to claim 1, characterized in that, The detection head adopts a decoupled design, adds a small target detection branch to process high-resolution feature maps, combines a scale-adaptive sample allocation algorithm to increase the proportion of positive samples of small targets, and uses the SIoU loss function to optimize bounding box regression.

6. The method according to claim 1, characterized in that, The method also includes a model inference optimization step, which uses a dynamic resolution mechanism and INT8 quantization compression to improve inference speed, reduce model size and energy consumption, and adapt to edge computing and real-time monitoring scenarios.

7. A waste detection system based on the above detection method, characterized in that, The system comprises a data acquisition and preprocessing module, an innovative feature extraction network module, a multi-scale feature fusion module, an attention-guided feature selection module, a decoupled multi-branch detection head module, an inference optimization and low-power deployment module, a post-processing and false alarm filtering module, and a visualization application interface module. These modules work together to achieve automatic detection, real-time monitoring, and efficient management of waste targets in urban environments.

8. The system according to claim 7, characterized in that, The decoupled multi-branch detection head module adopts a trajectory continuity verification and multimodal fusion strategy, combined with temporal information filtering and depth / thermal imaging data fusion, to reduce the false alarm rate and improve the detection reliability in complex environments.

9. The system according to claim 7, characterized in that, The visualization application interface module provides RESTful API, real-time visualization interface and automatic alarm and statistical report generation functions, supporting multi-platform integration and smart city management needs.

10. The system according to claims 7-9, characterized in that, The system is suitable for automatic waste identification, environmental monitoring, intelligent alarm and data analysis in large-scale urban environments, and features high-precision small target detection, low false alarm rate and high inference speed.

Images