An Automatic Detection Method for Solid Waste from Unmanned Aerial Vehicles Based on Feature Fusion

Abstract

This invention provides an automatic detection method for solid waste from drones based on feature fusion, comprising: acquiring images containing solid waste and generating an image set; using the image set as training samples for a solid waste CNN feature extraction network, and training the convolutional template parameters in the feature extraction network; training classification, proposal regression, and LFA mask networks using feature maps; segmenting morphological images; and introducing a histogram matching overlay analysis module to fuse the CNN and morphological solid waste detection results. This invention proposes an automatic detection method for solid waste from drones based on feature fusion, optimizing existing CNN structures, constructing a layer feature aggregation module based on multi-scale feature integration to fully utilize the image structure and deep semantic information of the target; designing an overlay analysis module to fuse CNN solid waste detection results, utilizing image gradient information and the spectral heterogeneity of solid waste to achieve accurate solid waste detection and edge adjustment; and constructing an automatic solid waste detection network through an improved feature fusion watershed mask CNN structure.

Description

Technical Field

[0001] This invention relates to the field of remote sensing information processing technology, and in particular to an automatic detection method for solid waste by unmanned aerial vehicles based on feature fusion. Background Technology

[0002] With the improvement of hardware performance (such as GPUs) and the continuous expansion of image processing technology and image object detection applications, convolutional neural networks (CNNs) have been widely used in many fields involving images and graphics, such as computer vision, artificial intelligence, and remote sensing. CNNs extract feature maps by convolving images (image synthesis). These feature maps are then used to form feature vectors for image classification (such as handwritten digit recognition and animal recognition) through fully connected layers. As application demands have increased, region convolutional neural networks have supplemented CNNs with object localization and segmentation capabilities. This means they not only identify the category of objects but also mark the relative position of specific objects in the image with bounding boxes, further extracting edge information of objects in the image. Currently, representative object detection methods in the field of deep learning include R-CNN, Faster R-CNN, Mask R-CNN, Cascade Mask R-CNN, YOLO, and DETR. Their common idea is to synthesize image information by using methods such as mean, maximum, minimum, and median of the R / G / B values ​​of each pixel, and then forming feature vectors through fully connected layers to train image classification networks.

[0003] In existing technical solutions, CNNs are used for image recognition in the fields of remote sensing, geosciences, and GIS.

[0004] (1) A large number of accurate real sample datasets are needed to drive the training and validation of CNNs. However, the boundaries of solid waste fragmentation and sharpening are difficult to annotate with high quality manually, which affects the accuracy of target edge information extraction.

[0005] (2) The spatial distribution of solid waste edges is relatively scattered and fragmented. If it is not magnified to a pixel-level scale for interpretation, it is difficult to extract sharpened boundary information. Therefore, in order to avoid misidentification and omissions, a lot of manpower and time are needed to check and verify each image in the manual visual interpretation process, which is inefficient and costly.

[0006] (3) After the target is processed by the CNN network through multiple layers of feature extraction and sampling, the sharpened and fragmented boundary information is excessively smoothed, which makes it impossible to effectively extract the solid waste edge, resulting in distorted target detection results.

[0007] Chinese invention patent application CN119672425A discloses a method and system for detecting solid waste. The method includes: obtaining solid waste samples of published categories, compositions, and contents, along with their corresponding Raman spectral data, from the Internet as a training set; using the training set to train an R-CNN-based target detection model; evaluating the model's generalization ability and performance through cross-validation after training; preparing solid waste samples and performing sample preprocessing after validation; scanning the samples with a Raman spectrometer after sample preprocessing to obtain Raman spectral data and then performing data preprocessing; and inputting the Raman spectral data into the validated R-CNN-based target detection model to output the detection result. The drawbacks of this method are: reliance on publicly available internet data may lead to inconsistent training set quality; different institutions use different Raman spectrometer parameters (such as laser wavelength and power) and sample preprocessing standards, introducing systematic errors into the spectral data; using AlexNet as the base network of R-CNN has significant limitations, as AlexNet's 5×5 large convolutional kernel (first layer) produces a low-resolution 55×55 feature map for a 227×227 input image, failing to effectively capture the fine features of the Raman spectrum image; no processing scheme was designed for multi-component mixtures. When the characteristic peaks of two wastes overlap, the model may misclassify them as a single substance; content prediction relies on a single regression head, without employing a multi-task learning framework, resulting in significant errors in component quantification.

[0008] Chinese invention patent application CN112990066A discloses a method and system for identifying solid waste from remote sensing images based on multi-strategy enhancement. The method includes cropping the original remote sensing image based on the input image and a preset overlap rate, saving the coordinates of the top-left corner of each cropped image block within the entire image; inputting the image blocks into a deep learning target detection network for solid waste identification, identifying the bounding boxes of the solid waste; the deep learning target detection network for solid waste identification employs a location-information-guided multi-strategy enhanced keypoint target recognition network; and filtering the detected bounding boxes to obtain the final identified bounding box result, which is then displayed in the original remote sensing image. The drawbacks of this method are: the use of an image cropping strategy with an overlap rate of 0.5 leads to the computation of a large number of overlapping regions, significantly increasing the amount of data processed and prolonging the processing time. For scenarios requiring rapid response in practical applications, this method may fail to meet real-time requirements due to computational redundancy. The network integrates multiple strategy enhancement modules (ASPP, PAN, attention mechanism, etc.) and location information-guided branches, resulting in a complex model structure and a large number of parameters. The training and inference processes require high-performance hardware, increasing deployment costs. The complex model may overfit on small-scale datasets, while solid waste annotation data is usually scarce, further limiting generalization ability. It relies on the precise matching of three key points: the top left corner, the bottom right corner, and the center point. However, solid waste in remote sensing images often presents irregular shapes or blurred boundaries (such as unclear edges of piles), making key point localization difficult. The bounding box selection relies only on an 80% overlap threshold and a highest score retention strategy, without considering the coexistence of multiple targets in complex scenarios. High-density distributed solid waste may be incorrectly merged due to partial overlap, leading to missed detections. Directly deleting low-score bounding boxes may ignore the real targets. Summary of the Invention

[0009] To address the aforementioned technical challenges, this invention proposes an automatic solid waste detection method for unmanned aerial vehicles (UAVs) based on feature fusion. It optimizes existing CNN structures by constructing a layer feature aggregation module based on multi-scale feature integration. This module fully utilizes the image structure and deep semantic information of targets in remote sensing image detection, improving target detection accuracy. Furthermore, it designs an overlay analysis module based on morphological watershed transformation and histogram matching to fuse CNN solid waste detection results, fully leveraging image gradient information and the spectral heterogeneity of solid waste to achieve accurate solid waste detection and edge adjustment. Finally, it constructs an automatic solid waste detection network using an improved feature fusion watershed mask CNN structure, improving the efficiency of solid waste interpretation, reducing manual intervention, and decreasing time and labor costs.

[0010] The purpose of this invention is to provide an automatic detection method for solid waste from unmanned aerial vehicles (UAVs) based on feature fusion, including acquiring images containing solid waste and generating an image set, and further including the following steps:

[0011] Step 1: Use the image set as training samples for the solid waste CNN feature extraction network to train the convolution template parameters in the feature extraction network;

[0012] Step 2: Train the classification, bounding box regression, and LFA masking networks using feature maps;

[0013] Step 3: Segment the morphological image;

[0014] Step 4: Introduce the histogram matching overlay analysis module, fuse the CNN solid waste segmentation results and morphological solid waste detection results to obtain the final model result and output it.

[0015] Preferably, the image containing solid waste is 600*600 pixels in size.

[0016] In any of the above schemes, step 1 includes the feature extraction network being non-linearly activated by an activation function, and the trained feature extraction network extracting feature maps of the input samples.

[0017] In any of the above solutions, step 2 preferably includes the following sub-steps:

[0018] Step 21: Use the training data and feature maps as training samples for the classifier, regression network, and mask network;

[0019] Step 22: Generate a preselection box with 15 possible combinations of pixels [48, 96, 192, 384, 768] pixels and aspect ratios [1:1, 2:1, 1:2];

[0020] Step 23: The preselected box approximates the target label box through a regression algorithm, serves as a suggestion box, and is further finely adjusted through the regression end to generate a predicted box;

[0021] Step 24: For the suggestion box containing the target object, resample the image to 128*128 pixels, compress all pixels into column vectors, and input them into the classification end composed of fully connected layers for solid waste classification;

[0022] Step 25: Construct a DSC using deformable convolutional and deconvolutional layers, introduce a multi-scale feature strategy to construct an LFA module, design an LFA mask network to accurately extract solid waste boundaries, generate each pixel category in the proposal box, and achieve pixel-level mask generation.

[0023] In any of the above solutions, step 3 preferably includes the following sub-steps:

[0024] Step 31: Expand the prediction box by N% and then crop it, where N is the pre-threshold for expansion;

[0025] Step 32: Perform binarization, erosion dilation, distance transform, and watershed segmentation on the cropped region to obtain a morphological segmentation result based entirely on image gradient;

[0026] Step 33: Fill the morphological image segmentation result to the same size as the original image.

[0027] In any of the above solutions, step 4 preferably includes the following sub-steps:

[0028] Step 41: Overlay the solid waste boundary graphic with the image extracted by CNN;

[0029] Step 42: Calculate the proportion of the solid waste intersection region in the image;

[0030] Step 43: Calculate the image histogram and measure the region similarity using the chi-square distance;

[0031] Step 44: Based on the overlay analysis, fuse the overlapping parts of the CNN segmentation results and the morphological segmentation results to obtain the final result.

[0032] In any of the above schemes, the overlay analysis is preferably defined as follows:

[0033]

[0034] in, A min and A max The minimum and maximum thresholds for the intersection area are given. D min and D max The minimum and maximum thresholds for the chi-square distance are given. Intersect This represents the proportion of the intersecting solid waste regions in the image. Dist To segment the image histogram, the chi-square distance is used. R m This is the solid waste area detected by CNN. R w Solid waste areas divided by morphological watersheds R f For the merged region, the symbol \ represents the difference between the two regions.

[0035] In any of the above schemes, the preferred aspect is that the proportion of the solid waste intersection area in the image is... Intersect The calculation formula is:

[0036]

[0037] in, COUNT () is a counting function, which is R w The number of pixels,n for R m and R w The number of pixels in the intersecting region.

[0038] In any of the above schemes, the chi-square distance is preferably defined as follows:

[0039]

[0040] in, H m and H w for R m and R w The corresponding image histogram, K This represents the number of intervals in the image histogram. H w ( k )and H m ( k ) is the corresponding histogram H w and H m The k Pixel frequency in each interval.

[0041] This invention proposes an automatic detection method for solid waste from drones based on feature fusion. On the basis of existing image target recognition algorithms, the method optimizes parameters for solid waste extraction and adds a multi-scale feature fusion and overlay analysis module for solid waste, thereby significantly improving the edge extraction accuracy of the automatic image target detection algorithm. Compared with existing image target detection and extraction algorithms for automatic solid waste detection, the average accuracy is improved by more than 20%, the F1 score is improved by 15%, and the edge accuracy of solid waste is significantly improved compared with manual visual interpretation. Attached Figure Description

[0042] Figure 1 This is a flowchart of a preferred embodiment of the UAV-based automatic solid waste detection method according to the present invention.

[0043] Figure 2 This is a flowchart of another preferred embodiment of the UAV-based automatic solid waste detection method according to the present invention.

[0044] Figure 3 This is a schematic diagram of an embodiment of the automatic detection method for unmanned aerial vehicle solid waste based on feature fusion according to the present invention, which extracts remote sensing image features through a trained CNN-like network.

[0045] Figure 4 This is a schematic diagram of an embodiment of the preselection box of the feature fusion-based automatic detection method for solid waste from unmanned aerial vehicles according to the present invention.

[0046] Figure 5 This is a schematic diagram of the structure of a layer feature aggregation mask network according to an embodiment of the automatic detection method for solid waste by unmanned aerial vehicles based on feature fusion according to the present invention.

[0047] Figure 6 This is a schematic diagram of an embodiment of the layer feature aggregation mask network for solid waste detection according to the feature fusion-based automatic detection method for unmanned aerial vehicles (UAVs) of the present invention.

[0048] Figure 7 This is a schematic diagram of an embodiment of the morphological segmentation image of the UAV-based automatic solid waste detection method according to the present invention.

[0049] Figure 8 This is a schematic diagram of an embodiment of the overlay analysis of the UAV solid waste automatic detection method based on feature fusion according to the present invention.

[0050] Figure 9 This is a schematic diagram of an embodiment of the overlay analysis-based boundary adjustment of the unmanned aerial vehicle (UAV) solid waste automatic detection method based on feature fusion according to the present invention. Detailed Implementation

[0051] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0052] Example 1

[0053] like Figure 1 As shown, an automatic detection method for solid waste by UAV based on feature fusion is described. Step 100 involves acquiring images containing solid waste and generating an image set, wherein the size of the images containing solid waste is 600*600 pixels.

[0054] Step 110 involves using the image set as training samples for a solid waste CNN feature extraction network, training the convolutional template parameters in the feature extraction network, and performing non-linear activation of the feature extraction network through an activation function. After training, the feature extraction network extracts feature maps from the input samples.

[0055] Perform step 120, training the classification, bounding box regression, and LFA masking networks using feature maps, including the following sub-steps:

[0056] Perform step 121, using the training data and feature maps as training samples for the classifier, regression network, and mask network.

[0057] Execute step 122 to generate 15 preselection boxes per pixel, with a width-to-length ratio of [1:1, 2:1, 1:2] and a total of [48, 96, 192, 384, 768] pixels.

[0058] In step 123, the preselected box approximates the target identifier box through a regression algorithm, serves as a suggestion box, and is further finely adjusted through the regression end to generate a predicted box.

[0059] In step 124, for the suggestion box containing the target object, the image is resampled to a size of 128*128 pixels, and all pixels are compressed into column vectors and input into the classification end composed of fully connected layers for solid waste classification.

[0060] In step 125, a DSC is constructed using deformable convolutional and deconvolutional layers, an LFA module is constructed by introducing a multi-scale feature strategy, an LFA mask network is designed to accurately extract solid waste boundaries, and each pixel category in the proposal box is generated to achieve pixel-level mask generation.

[0061] Perform step 130, segmenting the morphological image, including the following sub-steps:

[0062] Step 131 is executed, where the prediction box is expanded by N% and then cropped, where N is the expansion threshold. In this embodiment, N=5.

[0063] Step 132 is executed to perform binarization, erosion dilation, distance transformation and watershed segmentation on the cropped region to obtain a morphological segmentation result based entirely on image gradient.

[0064] Perform step 133 to fill the morphological image segmentation result to the same size as the original image.

[0065] Step 140 involves introducing a histogram matching overlay analysis module to fuse CNN and morphological solid waste detection results, including the following sub-steps:

[0066] Step 141 is executed, and the solid waste boundary graphic is superimposed on the image extracted by CNN.

[0067] Execute step 142 to calculate the proportion of the solid waste intersection area in the image. Intersect The calculation formula is:

[0068]

[0069] in, COUNT () is a counting function, which is R w The number of pixels, n for R m and R w The number of pixels in the intersecting region.

[0070] Perform step 143, calculate the image histogram, and measure region similarity using chi-square distance, where the chi-square distance is defined as follows:

[0071]

[0072] in, H m and H w for R m and R w The corresponding image histogram, K This represents the number of intervals in the image histogram. H w ( k )and H m ( k ) is the corresponding histogram H w and H m The k Pixel frequency in each interval.

[0073] Step 144 involves fusing the intersection of the CNN segmentation results and the morphological segmentation results based on overlay analysis to obtain the final result. Overlay analysis is defined as follows:

[0074]

[0075] in, A min and A max The minimum and maximum thresholds for the intersection area are given. D min and D max The minimum and maximum thresholds for the chi-square distance are given. Intersect This represents the proportion of the intersecting solid waste regions in the image. Dist To segment the image histogram, the chi-square distance is used. R m This is the solid waste area detected by CNN. R w Solid waste areas divided by morphological watersheds R f For the merged region, the symbol \ represents the difference between the two regions. .

[0076] Example 2

[0077] Target recognition in remote sensing images is significantly more complex than target recognition in general images due to the extremely complex background (image other than the target). Remote sensing images contain numerous irregular and heterogeneous patches. Furthermore, due to systematic errors in the remote sensing sensors themselves, they also exhibit distortion, salt-and-pepper noise, and other factors that affect the manual visual inspection of solid waste. In addition, solid waste is characterized by its scattered and fragmented nature, and sharpened, jagged boundaries, requiring pixel-by-pixel annotation at high resolution during manual interpretation—a labor-intensive, inefficient, and impractical task. The key to this invention is the efficient and automatic detection of solid waste in large-scale, high-precision remote sensing images through parameter optimization using deep learning methods and an LFA structure based on multi-scale feature fusion. This significantly improves the efficiency of solid waste detection and reduces the need for manual intervention.

[0078] Solid waste features numerous characteristics, and its high boundary complexity makes it difficult to obtain fine-grained sample datasets for deep learning training. Furthermore, existing CNN methods based on the image itself easily lose boundary information during convolution and downsampling, resulting in overly smoothed edges in detected solid waste. Therefore, the overlay analysis module that integrates CNN and morphological segmentation algorithms in this invention fully utilizes image gradients, semantics, and solid waste spectral information to achieve fine-grained solid waste extraction, which is another key aspect of this invention. This method represents an important extension of existing deep learning-based image detection techniques for geographic target detection in remote sensing.

[0079] The purpose of this invention is to develop a deep neural network model for automatic solid waste detection. This model enables automatic solid waste detection using UAV remote sensing imagery and related algorithms without human intervention. It provides an effective automated algorithm and computer program for the automatic extraction of solid waste. Based on existing image target recognition algorithms, this method optimizes parameters for solid waste extraction and incorporates a multi-scale feature fusion and overlay analysis module to significantly improve the edge extraction accuracy of the automatic image target detection algorithm. Compared with existing image target detection and extraction algorithms for automatic solid waste detection, the average accuracy is improved by over 20%, and the F1 score is improved by 15%. Compared with manual visual interpretation, it significantly improves the edge accuracy of solid waste.

[0080] This invention designs an LFA network based on a multi-scale feature fusion strategy and integrates it into existing image target detection algorithms. By fully utilizing the basic structure and deep semantic information of images, it significantly improves the accuracy of image recognition algorithms in remote sensing and geoscientific target identification, with an average accuracy improvement of over 20%. It also designs an overlay analysis module that integrates CNN and morphological target segmentation algorithms, which significantly improves the precision of solid waste boundary information compared to manual visual interpretation. Furthermore, it develops a watershed mask neural network for automatic detection of solid waste by UAVs based on feature fusion, which significantly improves the automation and efficiency of solid waste remote sensing interpretation and reduces the cost of manual interpretation.

[0081] Example 3

[0082] like Figure 2 As shown, an automatic detection method for solid waste from unmanned aerial vehicles (UAVs) based on feature fusion includes the following steps:

[0083] Step 1: Extract feature maps from remote sensing images. A collection of 600*600 pixel images containing solid waste is used as training samples for the solid waste CNN feature extraction network. The convolutional template parameters in the feature extraction network are trained. The feature extraction network undergoes non-linear activation using activation functions (such as ReLU, Sigmoid, Tanh). Figure 3 As shown, the feature extraction network extracts feature maps of the input samples after training.

[0084] Step 2: Train the classification, proposal regression, and LFA masking networks using feature maps. The training data and feature maps extracted in Step 1 are used as training samples for the classifier, regression network, and masking network. 15 pre-selected bounding boxes of [48, 96, 192, 384, 768] pixels with aspect ratios of [1:1, 2:1, 1:2] are generated pixel-by-pixel. These pre-selected boxes approximate the target bounding box using a regression algorithm, serving as proposals. Further fine-tuning is then performed on the regression side to generate predicted bounding boxes (e.g., ...). Figure 4 (As shown); for suggestion boxes containing target objects, the image is resampled to 128*128 pixels, and all pixels are compressed into column vectors. These are then input into a classification layer composed of fully connected layers for solid waste classification. A Deformable Convolutional Layer (DSC) is constructed using deformable convolutional and deconvolutional layers, and a multi-scale feature strategy is introduced to build an LFA module, such as... Figure 5 The design shown is an LFA masking network that accurately extracts solid waste boundaries, generating a category for each pixel in the suggestion box, such as... Figure 6 The diagram illustrates pixel-level mask generation.

[0085] Step 3: Morphological Image Segmentation. Following Step 2, solid waste classification, bounding box generation, and mask generation were achieved. To include all solid waste pixels within the bounding box, the box was enlarged and cropped proportionally. The cropped region underwent binarization, erosion and dilation, distance transformation, and watershed segmentation to obtain a morphological segmentation result entirely based on image gradients. The morphological image segmentation result was then filled to the same size as the original image. The segmentation result is shown below. Figure 7 As shown.

[0086] Step 4: Introduce a histogram matching overlay analysis module to fuse the CNN and morphological solid waste detection results. The more complex and detailed solid waste boundary graphics obtained in Step 3 are then overlaid with the images extracted by the CNN. For example... Figure 8 As shown, the overlay analysis is defined as follows:

[0087]

[0088] in, A min and A max It is the intersection area threshold. D min and D max It is the chi-square distance threshold. Intersect It is the ratio of the overlapping areas of the segmentation results. Dist It is the chi-square distance of the histogram of the segmented image. R m This is the solid waste area detected by CNN. R w It is a solid waste area divided by a morphological watershed. R f It represents the merged region. The symbol "\" represents the difference between the two regions. .

[0089] The proportion of the intersecting solid waste areas in the image is calculated using the following formula:

[0090]

[0091] in, COUNT () is a counting function. N yes R w The number of pixels, n This is for R m and R w The number of pixels in the intersecting region.

[0092] To fully utilize the spectral characteristics of ground features to aid in overlay analysis, image histograms are calculated, and chi-square distance is used to measure regional similarity, thereby determining whether fusion is appropriate. The chi-square distance is defined as follows:

[0093]

[0094] in, H m and H w yes R m and R w The corresponding image histogram is the number of intervals in the image histogram. H w ( k )and H m ( k ) is the corresponding histogram H wand H m The k Pixel frequency in each interval.

[0095] like Figure 9 As shown, the CNN solid waste segmentation results and the morphological solid waste segmentation results are fused based on overlay analysis to achieve fine adjustment of solid waste boundaries, resulting in a final solid waste segmentation result with accurate fine-grained boundaries.

[0096] This application optimizes the existing CNN structure by constructing a layer feature aggregation module based on multi-scale feature integration, which fully utilizes the image structure and deep semantic information of the target in remote sensing image target detection, thereby improving the target detection accuracy.

[0097] This application designs an overlay analysis module based on morphological watershed transformation and histogram matching, which integrates CNN solid waste detection results and makes full use of image gradient information and solid waste spectral heterogeneity to achieve accurate solid waste detection and edge adjustment.

[0098] This application constructs an automatic solid waste detection network using an improved feature fusion watershed mask CNN structure, thereby improving the efficiency of solid waste interpretation, reducing human intervention, and decreasing time and labor costs.

[0099] To better understand this invention, specific embodiments have been described in detail above, but these are not intended to limit the invention. Any simple modifications made to the above embodiments based on the technical essence of this invention still fall within the scope of this invention. Each embodiment in this specification focuses on its differences from other embodiments; similar or identical parts between embodiments can be referred to mutually. For system embodiments, since they basically correspond to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

Claims

1. A method for automatic detection of solid waste by unmanned aerial vehicles (UAVs) based on feature fusion, comprising acquiring images containing solid waste and generating an image set, characterized in that, It also includes the following steps: Step 1: Use the image set as training samples for the solid waste CNN feature extraction network to train the convolution template parameters in the feature extraction network; The images containing solid waste are remote sensing images; Step 2: Train the classification, bounding box regression, and LFA masking networks using feature maps; Step 3: Segment the morphological image; Step 4: Introduce a histogram matching overlay analysis module to fuse CNN solid waste segmentation results and morphological solid waste detection results, thereby achieving fine-grained adjustment of solid waste boundaries. This results in the final solid waste segmentation result with accurate, fine-grained boundaries, and is then output. This includes the following sub-steps: Step 41: Overlay the solid waste boundary graphic with the image extracted by CNN; Step 42: Calculate the proportion of the solid waste intersection region in the image; Step 43: Calculate the image histogram and measure the region similarity using the chi-square distance; Step 44: Based on the overlay analysis, fuse the intersection of the CNN segmentation results and the morphological segmentation results to obtain the final result. The overlay analysis is defined as follows: ， in, A min and A max The minimum and maximum thresholds for the intersection area, D min and D max The minimum and maximum thresholds for the chi-square distance are given. Intersect This represents the proportion of the intersecting solid waste regions in the image. Dist To segment the image histogram, the chi-square distance is used. R m This is the solid waste area detected by CNN. R w Solid waste areas divided by morphological watersheds R f For the merged region, the symbol \ represents the difference between the two regions; The proportion of the solid waste intersection area in the image Intersect The calculation formula is: ， in, COUNT () is a counting function, which is R w The number of pixels, n for R m and R w The number of pixels in the intersecting region.

2. The automatic detection method for solid waste from unmanned aerial vehicles based on feature fusion as described in claim 1, characterized in that, The image containing solid waste is 600*600 pixels in size.

3. The automatic detection method for solid waste from unmanned aerial vehicles based on feature fusion as described in claim 2, characterized in that, Step 2 includes the following sub-steps: Step 21: Use the training data and feature maps as training samples for the classifier, regression network, and mask network; Step 22: Generate a preselection box with 15 possible combinations of pixels [48, 96, 192, 384, 768] pixels and aspect ratios [1:1, 2:1, 1:2]; Step 23: The preselected box approximates the target label box through a regression algorithm, serves as a suggestion box, and is further finely adjusted through the regression end to generate a predicted box; Step 24: For the suggestion box containing the target object, resample the image to 128*128 pixels, compress all pixels into column vectors, and input them into the classification end composed of fully connected layers for solid waste classification; Step 25: Construct a DSC using deformable convolutional and deconvolutional layers, introduce a multi-scale feature strategy to construct an LFA module, design an LFA mask network to accurately extract solid waste boundaries, generate each pixel category in the proposal box, and achieve pixel-level mask generation.

4. The automatic detection method for solid waste from unmanned aerial vehicles based on feature fusion as described in claim 3, characterized in that, Step 3 includes the following sub-steps: Step 31: Expand the prediction box by N% and then crop it, where N is the pre-threshold for expansion; Step 32: Perform binarization, erosion dilation, distance transform, and watershed segmentation on the cropped region to obtain a morphological segmentation result based entirely on image gradient; Step 33: Fill the morphological image segmentation result to the same size as the original image.

5. The automatic detection method for solid waste from unmanned aerial vehicles based on feature fusion as described in claim 4, characterized in that, The definition of the chi-square distance is... ， in, H m and H w for R m and R w The corresponding image histogram, K This represents the number of intervals in the image histogram. H w ( k )and H m ( k ) is the corresponding histogram H w and H m The k Pixel frequency in each interval.

Images