Lightweight recognition method and system for cubilose fruits and flowers in low-light environment based on LLDF-YOLO11

Abstract

The invention discloses an LLDF-YOLO11-based lightweight recognition method for cubilose fruits and flowers in a low-illumination environment. The method comprises the following steps: acquiring a to-be-recognized low-illumination cubilose fruit and flower image and a trained LLDF-YOLO11 model; inputting a to-be-identified low-illumination cubilose fruit and flower image into the backbone network; performing illumination calibration through an SCInet module of the backbone network to enhance the image quality, and obtaining the cubilose fruit and flower image after illumination calibration; the bird's nest fruit and flower image after illumination calibration passes through a hierarchical structure composed of a conv convolutional layer, a plurality of stacked "ADown downsampling modules + C3k2 modules", an SPPF module and a C2PSALSKA module, so that a multilayer feature map of the bird's nest fruit and flower image after illumination calibration is extracted; inputting the multi-layer feature map output by the backbone network into an ICCFM neck network, performing multi-scale fusion and enhancement on features from different levels through the ICCFM neck network to integrate context information, and generating a multi-scale cubilose fruit and flower fusion feature map; and sending the obtained multi-scale cubilose fruit and flower feature map into a corresponding detection head, and predicting position coordinates and category confidence of cubilose fruits and flowers in the map. According to the method, accurate and efficient recognition of cubilose fruits and flowers in a low-light environment is realized, and technical support is provided for automatic pollination.

Description

Technical Field

[0001] This invention relates to a lightweight identification method and system for bird's nests, fruits, and flowers in low-light environments based on LLDF-YOLO11, belonging to the fields of image processing and intelligent agriculture. Background Technology

[0002] Dragon fruit (also known as yellow-skinned pitaya or dragon fruit) is a tropical fruit with extremely high nutritional value, and it is already being cultivated on a small scale in Hainan, Yunnan, Guangdong and other regions of my country. Its flowers only open briefly at night and wither the following morning, resulting in a very low natural pollination rate. Artificial pollination is necessary to meet the needs of commercial production. However, manual pollination is labor-intensive and costly, making automated pollination an inevitable trend in the industry's development.

[0003] With the rapid development of deep learning, neural network-based object detection algorithms have become the mainstream technology in the field of object detection. Deep learning-based object detection techniques can generally be divided into two categories: one-stage and two-stage. Two-stage object detection algorithms first use a candidate region generation strategy to generate candidate regions that may contain the target, and then perform target location regression and category prediction on these candidate regions. Representative methods include Faster R-CNN and FPN. Single-stage object detection does not require extracting candidate regions; it only needs one forward inference to obtain the target category and location. Representative methods include SSD and YOLO. Two-stage object detection uses a coarse-to-fine processing logic, resulting in high detection accuracy but slower speed, and is rarely used in engineering practice. Conversely, single-stage object detection usually has slightly lower accuracy but better real-time performance. YOLO, due to its fast inference efficiency and small model size, has become the preferred object detection algorithm at present.

[0004] However, the core prerequisite for automated pollination is accurate identification of flowers to be pollinated, but flower recognition for bird's nest fruit faces multiple challenges: pollination operations need to be carried out at night in low light, resulting in images with high noise, low contrast, and blurred details; flowers are easily obscured by branches, densely overlapping, and mixed with buds and withered flowers. Existing mainstream target detection models are mostly designed for daytime environments with sufficient light, resulting in low detection accuracy in complex low-light scenarios, and the models have a large number of parameters and high computational complexity, making them difficult to deploy on agricultural edge devices. Summary of the Invention

[0005] This invention provides a lightweight identification method and system for bird's nest fruit flowers in low-light environments based on LLDF-YOLO11, enabling accurate and efficient identification of bird's nest fruit flowers in low-light environments and providing technical support for automated pollination.

[0006] The technical solution of this invention is:

[0007] According to a first aspect of the present invention, a lightweight identification method for bird's nest flowers in low-light environments based on LLDF-YOLO11 is provided.

[0008] Step 1: Obtain the low-light image of the bird's nest fruit flower to be identified and the trained LLDF-YOLO11 model; the LLDF-YOLO11 model consists of three parts: the backbone network, the ICCFM neck network, and the detection head;

[0009] Step 2: Input the low-light bird's nest and fruit flower image to be identified into the backbone network: The SCI_net module of the backbone network is used to perform illumination calibration to enhance the image quality and obtain the illumination-calibrated bird's nest and fruit flower image; The illumination-calibrated bird's nest and fruit flower image is passed through a hierarchical structure consisting of conv convolutional layers, multiple stacked "ADown downsampling module + C3k2 module", SPPF module, and C2PSA_LSKA module to extract multi-layer feature maps of the illumination-calibrated bird's nest and fruit flower image;

[0010] Step 3: Input the multi-layer feature map output from the backbone network in Step 2 into the ICCFM neck network. The ICCFM neck network performs multi-scale fusion and enhancement on features from different levels to integrate contextual information and generate a multi-scale bird's nest, fruit and flower fusion feature map.

[0011] Step 4: Input the multi-scale bird's nest fruit flower feature map obtained in Step 3 into the corresponding detection head to predict the position coordinates and category confidence of the bird's nest fruit flower in the map.

[0012] Furthermore, the backbone network of the LLDF-YOLO11 model is specifically as follows: The SCI_net module is fused before layer 0 of the YOLOv11n basic framework; the standard convolutions in layers 1, 3, 5, and 7 of the traditional YOLOv11n backbone network are replaced with an ADown downsampling module. The ADown downsampling module includes an average pooling layer, channel splitting operation, a 3×3 convolutional layer, a max pooling layer, a 1×1 convolutional layer, and a concatenation layer; the input feature map is split into two paths after the average pooling layer extracts global information, and then split into two paths after the channel splitting operation. One path extracts local details through a 3×3 convolutional layer, and the other path captures salient features through a max pooling layer and then compresses the channels through a 1×1 convolutional layer. Finally, the two feature paths are concatenated and output; the C2PSA module in layer 10 of the traditional YOLOv11n backbone network is replaced with a C2PSA_LSKA module; the C2PSA_LSKA module combines the LSKA attention mechanism with the parallel spatial attention mechanism of C2PSA.

[0013] Furthermore, the SCI_net module includes an illumination estimation module and a self-calibration module, which are connected in a cascaded manner. The training and testing processes of the SCI_net module are as follows:

[0014] During the training of the SCI_net module, the output of the element-wise division operation between the original low-light bird's nest and fruit flower image and the current illumination image is input into the parameterization operator to generate a calibration term. The original low-light bird's nest and fruit flower image is then element-wise added to the calibration term to obtain the calibrated output, which serves as the input to the illumination estimation module. The input to the illumination estimation module is then processed by a convolution module to generate a residual term. The current illumination image is then element-wise added to the residual term to obtain the updated illumination image for the next stage. This iterative process continues until the next iteration. The process involves several stages, ultimately resulting in a trained illumination map.

[0015] During the testing of the SCI_net module, the low-light bird's nest and fruit flower image to be tested is used as input to the illumination estimation module to generate an illumination map to be enhanced. The original low-light bird's nest and fruit flower image to be tested is divided element-wise with the illumination map to be enhanced to obtain the final enhanced output.

[0016] Furthermore, the ADown downsampling module specifically works as follows: the initial feature map is fed into the ADown downsampling module, where global information is first extracted and the dimensionality is reduced through an average pooling layer; then, the feature map is split into two branches along the channel dimension: one branch extracts local detail features through a 3×3 convolution; the other branch captures significant features through a max pooling layer, compresses the channels through a 1×1 convolution, and is given non-linear properties; finally, the features from the two branches are concatenated and fused along the channel dimension to obtain the downsampled feature map.

[0017] Furthermore, the ICCFM neck network of the LLDF-YOLO11 model adopts an ICCFM structure, specifically as follows:

[0018] The multi-layer feature maps output by the backbone network are first preliminarily adjusted in terms of channel dimensions and features through convolutional layers to obtain dimension-matched feature maps to be fused. , , ;

[0019] Future deep feature maps of autonomous backbone networks The resolution is increased by using the Upsample module, making it comparable to the feature maps of the layers in the autonomous backbone network. The two maps, with consistent scale, are concatenated and fused, and then refined by the C3k2 module to generate an intermediate feature map. Subsequently, this intermediate feature map is further amplified in resolution by passing through convolutional layers and an upsample module, and then combined with feature maps from shallow layers of the backbone network. The splicing is completed along the channel dimension, and finally the features are integrated through the C3k2 module to generate a top-level small-scale feature map containing rich semantic and spatial information.

[0020] The generated top-level small-scale feature map is input into the ADown module for downsampling, concatenated with the feature map of the intermediate feature map after convolution, and then fused by the C3k2 module to obtain the mid-level mesoscale feature map; next, this mid-level mesoscale feature map is fed into another ADown module for downsampling, and then combined with the generated... The images are stitched together and processed by the C3k2 module to obtain the underlying large-scale feature map of the large-scale receptive field.

[0021] Furthermore, the training of the LLDF-YOLO11 model specifically involves: acquiring and filtering low-light images of bird's nest fruit; enhancing and cropping the filtered bird's nest fruit images to construct a dataset; and training and validating the LLDF-YOLO11 model based on the training and validation sets in the dataset to obtain a well-trained LLDF-YOLO11 model with optimized accuracy and efficiency metrics.

[0022] Furthermore, the collaborative optimization of accuracy-related indicators and efficiency-related indicators specifically involves: the accuracy-related indicator value reaching the preset threshold of the accuracy-related indicator, and the lightweight evaluation indicator value being lower than the lightweight evaluation indicator threshold.

[0023] Furthermore, the accuracy-related indicator values ​​reaching the preset thresholds include: an accuracy value above a first preset threshold, a recall rate above a second preset threshold, and an average precision value above a third preset threshold; the lightweight evaluation indicator values ​​below the lightweight evaluation indicator thresholds include: the number of model parameters below a fourth preset threshold.

[0024] Furthermore, the acquisition of the low-light bird's nest fruit images specifically refers to the acquisition periods being 21:00-22:00 and 7:00-8:00.

[0025] According to a second aspect of the present invention, a lightweight identification system for bird's nests, fruits, and flowers in low-light environments based on LLDF-YOLO11 is provided, comprising a module of the lightweight identification method for bird's nests, fruits, and flowers in low-light environments based on LLDF-YOLO11 as described above.

[0026] The beneficial effects of this invention are:

[0027] 1. This invention improves recognition accuracy and robustness in low-light environments: By introducing the SCI_net module at the front end of the backbone network, this invention performs illumination calibration and enhancement on images of bird's nests and flowers under low-light conditions, effectively solving problems such as high image noise, low contrast, and blurred details caused by insufficient light during nighttime pollination operations. Compared with the baseline model YOLOv11n, the LLDF-YOLO11 model of this invention achieves higher accuracy (P) and mean average precision (mAP). 0.5-0.95 The accuracy and performance of the model improved by 1.9 and 4.4 percentage points respectively, significantly enhancing the model's ability to identify targets in complex low-light scenarios.

[0028] 2. This invention achieves lightweight model design, facilitating deployment on edge devices: By employing the ICCFM neck network structure, the ADown downsampling module, and the C2PSA_LSKA attention module, this invention significantly optimizes the model structure while improving model performance. The final model has only 1.40M parameters and 5.1G of computational resources, a 45.7% reduction in parameters compared to the baseline model YOLOv11n. This makes the model less demanding on hardware resources, easier to deploy on agricultural edge computing devices, and provides an efficient and low-cost technical solution for automated pollination.

[0029] 3. This invention enhances multi-scale feature fusion capabilities, effectively mitigating the loss of detailed information: Addressing issues such as dense overlap and branch occlusion in bird's nest fruit flowers, the ADown downsampling module and ICCFM neck network employed in this invention effectively fuse feature information at different scales, mitigating the loss of subtle features that may occur with traditional downsampling operations. This enables the model to better capture and identify flowers of different sizes and shapes, reducing the false negative and false positive rates in dense, occluded scenes, and improving detection accuracy. Attached Figure Description

[0030] Figure 1 This is a network structure diagram of the LLDF-YOLO11 model of the present invention.

[0031] Figure 2 This is a schematic diagram of the SCI_net framework of the present invention.

[0032] Figure 3 This is a structural diagram of the ADown module of the present invention.

[0033] Figure 4 The diagram shows the structure of LSKA and C2PSA_LSKA of the present invention; (a) represents the structure of LSKA, (b) represents the structure of C2PSA_LSKA, and (c) represents the structure of PSABlock.

[0034] Figure 5 This is the dataset of bird's nest and fruit flower images used in this embodiment.

[0035] Figure 6 Comparison of detection results using the embodiments of this invention and the YOLO11n model for the bird's nest and fruit flower image dataset. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be arbitrarily combined with each other.

[0037] Example 1: As Figures 1-6 As shown, according to a first aspect of the present invention, a lightweight identification method for bird's nest fruit flowers in low-light environments based on LLDF-YOLO11 (Lightweight Low-light Dragon Fruit Flower-YOLO11) is provided, comprising:

[0038] Step 1: Obtain the low-light image of the bird's nest fruit flower to be identified and the trained LLDF-YOLO11 model; the LLDF-YOLO11 model consists of three parts: the backbone network, the ICCFM neck network, and the head.

[0039] Step 2: Input the low-light bird's nest and fruit flower image to be identified into the backbone network: The SCI_net module of the backbone network is used to perform illumination calibration to enhance the image quality and obtain the illumination-calibrated bird's nest and fruit flower image; The illumination-calibrated bird's nest and fruit flower image is passed through a hierarchical structure consisting of conv convolutional layers, multiple stacked "ADown downsampling module + C3k2 module", SPPF module, and C2PSA_LSKA module to extract multi-layer feature maps of the illumination-calibrated bird's nest and fruit flower image;

[0040] Step 3: Input the multi-layer feature map output from the backbone network in Step 2 into the ICCFM neck network. The ICCFM neck network performs multi-scale fusion and enhancement on features from different levels to integrate contextual information and generate a multi-scale bird's nest, fruit and flower fusion feature map that is robust to targets of different sizes.

[0041] Step 4: Input the multi-scale bird's nest fruit flower feature map obtained in Step 3 into the corresponding detection head to predict the position coordinates and category confidence of the bird's nest fruit flower in the map.

[0042] Furthermore, the backbone network of the LLDF-YOLO11 model is specifically as follows: an SCI_net module is fused before layer 0 of the YOLOv11n basic framework; the standard convolutions in layers 1, 3, 5, and 7 of the traditional YOLOv11n backbone network are replaced with an ADown downsampling module. The ADown downsampling module includes an average pooling layer, channel splitting operation, a 3×3 convolutional layer, a max pooling layer, a 1×1 convolutional layer, and a concatenation layer; after the input feature map is processed by the average pooling layer to extract global information, it is split into two paths by the channel splitting operation. One path is processed by a 3×3 convolutional layer to extract local details, and the other path is processed by a max pooling layer to capture significant features and then processed by a 1×1 convolutional layer to compress channels. Finally, the two feature paths are concatenated and output. The ADown downsampling module can effectively alleviate the loss of details caused by over-downsampling while reducing the feature map size through multi-scale feature fusion. The C2PSA module in the 10th layer of the traditional YOLOv11n backbone network is replaced with the C2PSA_LSKA module; the C2PSA_LSKA module combines the LSKA attention mechanism with the parallel spatial attention mechanism of C2PSA.

[0043] like Figure 2 As shown, the SCI_net module is based on the Self-Calibrated Illumination (SCI) learning framework. The SCI_net module mainly consists of two core modules: an Illumination Estimation module and a Self-Calibrated Module. These two modules are cascaded together. The training and testing processes for the SCI_net module are as follows:

[0044] During the training of the SCI_net module, the original low-light images of bird's nests and flowers were used. Compared with the current lighting map Performing element-wise division, we get , Next, enter the parameterized operator. In the process, generate calibration items. ( (Represents a network structure consisting of N 3×3 convolutional layers + ReLU activation function), original low-light image of bird's nest and fruit flowers. With calibration items Perform element-wise addition to obtain the calibrated output. As input to the illumination estimation module; input to the illumination estimation module via convolution module ( (This represents a convolutional module consisting of three 3×3 convolutional layers plus a ReLU activation function), generating residual terms. The current lighting map With residuals Perform element-wise addition to obtain the updated lighting map for the next stage. This iterative process continues until the [number]th [number]. This process involves several stages, ultimately resulting in a fully trained lighting map. During the zero-stage iteration, the original low-light image of the bird's nest and fruit flowers was used. As the initial lighting diagram The first-stage updated lighting map is obtained through the SCI_net module. .

[0045] During the testing of the SCI_net module, the low-light image of bird's nest and fruit flowers to be tested was used. The original low-light image of the bird's nest flower to be tested is used as input to generate an illumination map to be enhanced, which is then used as input to the illumination estimation module. Perform element-wise division with the lighting map to be enhanced to obtain the final enhanced output. .

[0046] Based on the Self-Calibrated Illumination (SCI) framework, an SCI_net module was constructed and embedded into the 0th layer of the backbone network of the YOLOv11n object detection network. The aim is to perform accurate illumination correction on bird's nest fruit images in low-light environments, thereby effectively improving the model's target recognition accuracy and robustness in this scenario.

[0047] It should be noted that the traditional YOLOv11n backbone network includes multiple standard convolutional layers, multiple C3k2 layers, one SPPF layer, and one C2PSA layer. These layers are connected serially and numbered sequentially from layer 0 to layer 10. Layers 0, 1, 3, 5, and 7 are standard convolutional layers; layers 2, 4, 6, and 8 are C3k2 layers; layer 9 is an SPPF layer; and layer 10 is a C2PSA layer. The backbone network of this invention, based on this, uses SCI_net as layer 0 to perform illumination calibration on the input image, outputting a clearly defined, illumination-calibrated image of the bird's nest flower.

[0048] like Figure 3As shown, the specific working process of the ADown downsampling module is as follows: The initial feature map is fed into the ADown downsampling module. First, global information is extracted and the dimension is reduced through the average pooling layer (AvgPool2d). Then, the feature map is split into two branches along the channel dimension: (1) one branch extracts local detail features through 3×3 convolution; (2) the other branch captures significant features through the max pooling layer (MaxPool2d), and then compresses the channels through 1×1 convolution and gives it non-linear characteristics. Finally, the features of the two branches are concatenated (Concat) along the channel dimension to obtain the downsampled feature map. This multi-scale feature fusion strategy effectively alleviates the loss of details caused by over-downsampling, and its expression is as follows:

[0049] ;

[0050] in, Indicates one input path, This is the input for another path.

[0051] like Figure 4 As shown, the C2PSA_LSKA module is composed of the parallel spatial attention mechanism of C2PSA and the attention mechanism of LSKA (Large Separable Kernel Attention).

[0052] LSKA will be a two-dimensional Depth convolution decomposes into two one-dimensional depth convolutions (one horizontal and one vertical). This operation can significantly reduce computation and the number of parameters, and is used to capture local contextual information. LSKA transforms a two-dimensional... The deep-dilated convolution is decomposed into two one-dimensional deep-dilated convolutions, which aim to expand the receptive field without increasing computational cost, thereby capturing long-range dependencies.

[0053] The LSKA attention mechanism is as follows:

[0054] (1) The input feature map is processed by two depthwise convolutions and two depthwise dilated convolutions:

[0055]

[0056]

[0057] in, It is the output after two depthwise convolutions (DW). It is the output after two depthwise dilated convolutions (DW-D-Conv). For the input feature map, Indicates the void ratio, Let be the side length of the convolution kernel. Indicates channel corresponding Size weight, This represents summing the height H and width W dimensions of the feature map. Represents convolution operation. This represents the floor function.

[0058] (2) Through a 1×1 convolution To fuse the channel information from the outputs of two deep-dilated convolutions and generate an attention map:

[0059] ;

[0060] (3) Multiply the generated attention map element-wise with the original input feature map to reweight the features and highlight important regions:

[0061] ;

[0062] C2PSA alone has the following drawbacks: limited ability to model long-distance dependencies; insufficient aggregation of contextual information; and reduced sensitivity to spatially dispersed features. Therefore, fusing C2PSA and LSKA into C2PSA_LSKA can effectively improve these shortcomings, thereby achieving a synergistic balance between fine-grained local details and global contextual dependencies.

[0063] Furthermore, the neck network of the LLDF-YOLO11 model adopts an ICCFM structure. The ICCFM structure replaces the standard convolution before the concat operation in the CCFM structure with an ADown downsampling module. This invention replaces the original neck network of the traditional YOLOv11n with an ICCFM (Improve Cross-Scale Feature Fusion Module) structure. The ICCFM structure integrates features from adjacent scales through fusion blocks, replaces the standard convolution before concat with an ADown module, and finally compresses the number of channels using 1×1 convolutions to control computational load and enhance multi-scale feature fusion capabilities.

[0064] The specific working process of ICCFM is as follows:

[0065] (1) Feature input and dimensionality adjustment: multi-layer feature maps output by the backbone network ( First, a convolutional layer (Conv) is used to perform preliminary adjustments to the channel dimensions and features, resulting in a dimension-matched feature map to be fused. ( , , This process can be represented as:

[0066] ;

[0067] (2) Cross-scale upsampling feature interaction: feature maps to be fused ( , , Bottom-up feature fusion will be performed. First, the feature maps of the deeper layers of the future autonomous backbone network will be fused. (The bottom branch in the diagram) improves the resolution through the Upsample module, making it compatible with the feature maps of the layers in the autonomous backbone network. (The middle branch in the image) has the same scale, and the two are concatenated and fused. The C3k2 module then refines the data to generate an intermediate feature map. Subsequently, this intermediate feature map is further amplified in resolution by passing through a convolutional layer (Conv) and an upsample module, and then combined with feature maps from the shallow layers of the backbone network. (The top branch in the figure) is concatted along the channel dimension, and finally the C3k2 module is used for feature integration to generate a top-level small-scale feature map containing rich semantic and spatial information.

[0068] (3) ADown downsampling adaptation and deep fusion: First, the top-level small-scale feature map generated in process (2) is input into the ADown module for downsampling, and concatenated with the feature map after convolution of the intermediate feature map generated in process (2). After fusion by the C3k2 module, a mid-level mesoscale feature map is obtained. Then, the mid-level mesoscale feature map is sent into an ADown module for downsampling again, and then concatenated with the feature map generated in process (1). The layers are concatted and then processed by the C3k2 module to obtain the underlying large-scale feature map of the large-scale receptive field.

[0069] (4) Feature extraction and detection adaptation: The three-layer feature maps after deep fusion are sent to the detection head to realize multi-scale target detection.

[0070] Small object detection stream: The top-level small-scale feature map (highest resolution) retains rich detail information and is directly input into the first Detect head to detect tiny features in the image (such as flower bud tips).

[0071] Mid-scale target detection flow: Mid-level mesoscale feature maps combine semantics and details, and are input to a second Detect head for detecting mid-scale targets (such as complete flowers).

[0072] Large target detection flow: The bottom-level large-scale feature map (lowest resolution) is downsampled multiple times to have the largest receptive field. It is input into the third Detect head to detect large-scale targets (such as patches of plants or large objects with complex backgrounds).

[0073] Furthermore, the training of the LLDF-YOLO11 model specifically involves: acquiring and filtering images of bird's nest fruit; enhancing and cropping the filtered images to construct a dataset; dividing the dataset into a training set, a validation set, and a test set according to a preset ratio; and training and validating the LLDF-YOLO11 model using the training and validation sets to obtain a well-trained LLDF-YOLO11 model with optimized accuracy and efficiency metrics. The test set is used for testing.

[0074] Furthermore, the collaborative optimization of accuracy-related indicators and efficiency-related indicators specifically involves: the accuracy-related indicator value reaching the preset threshold of the accuracy-related indicator, and the lightweight evaluation indicator value being lower than the lightweight evaluation indicator threshold.

[0075] Furthermore, the accuracy-related metric values ​​reaching the preset thresholds include: an accuracy value above the first preset threshold (e.g., above 97%), a recall rate above the second preset threshold (e.g., above 97%), and an average precision value above the third preset threshold (e.g., above 80%); the lightweight evaluation metric values ​​below the lightweight evaluation metric thresholds include: a model parameter count below the fourth preset threshold (e.g., below 1.5M), and FLOPs below the fifth preset threshold (e.g., below 5.5G).

[0076] Furthermore, the data enhancement operations include horizontal flipping, vertical flipping, random adjustment of contrast, random adjustment of brightness, and random adjustment of saturation; the cropping operation is a 1:1 center cropping (maintaining height and center alignment).

[0077] According to a second aspect of the present invention, a lightweight recognition system for bird's nests and fruit flowers in low-light environments based on LLDF-YOLO11 is provided, comprising modules of the lightweight recognition method for bird's nests and fruit flowers in low-light environments based on LLDF-YOLO11 described above; specifically including: an image acquisition module for acquiring images of bird's nests and fruit flowers to be recognized under low light conditions; and a prediction module for acquiring a trained LLDF-YOLO11 model; the LLDF-YOLO11 model consists of a backbone network, an ICCFM neck network, and a head; the image of the bird's nest and fruit flowers to be recognized under low light conditions is input into the backbone network; and light calibration is performed through the SCI_net module of the backbone network to... Image quality is enhanced to obtain a light-calibrated image of bird's nest and fruit flowers. This image is then processed through a hierarchical structure consisting of convolutional layers, multiple stacked "ADown downsampling module + C3k2 module," SPPF module, and C2PSA_LSKA module to extract multi-layer feature maps. The multi-layer feature maps output from the backbone network are input into the ICCFM neck network. The ICCFM neck network performs multi-scale fusion and enhancement on features from different layers to integrate contextual information and generate a multi-scale fused feature map of bird's nest and fruit flowers that is robust to targets of different sizes. The obtained multi-scale feature map is then fed into the corresponding detection head to predict the position coordinates and category confidence of the bird's nest and fruit flowers in the image. For details not elaborated on in the above modules, please refer to the relevant descriptions in this embodiment.

[0078] Example 2: The implementation process of the present invention is described below with reference to experiments:

[0079] Step 1: Image Acquisition and Data Augmentation

[0080] The images were taken at the Saimaba Bird's Nest Fruit Planting Base in Lujiang Town, Baoshan City, Yunnan Province, using an iPhone 14 Pro. The capture times were 21:00-22:00 (night) and 7:00-8:00 (early morning). The capture angles included eye-level and overhead views, with a distance of 50cm-100cm from the flowers (60±5cm in this embodiment of the invention). The original image resolution was 1280×720.

[0081] Dataset creation process:

[0082] The 2071 collected images were quality-screened, blurry and duplicate samples were removed, and 710 basic dataset images were retained;

[0083] Data augmentation operations were performed on the basic dataset, including horizontal flipping, vertical flipping, random adjustment of contrast, random adjustment of brightness, and random adjustment of saturation, expanding it to 4260 images.

[0084] Crops all images to 720×720 pixel resolution at a 1:1 ratio (keeping height and center alignment).

[0085] Manual labeling was performed using Labelimg software. The categories included bud, blooming flower, and wilted flower, and the labeling format was the minimum bounding rectangle.

[0086] The dataset is divided into a training set (3408 images), a validation set (426 images), and a test set (426 images) in an 8:1:1 ratio.

[0087] Step 2: Construct the LLDF-YOLO11 model;

[0088] Step 3: Train the LLDF-YOLO11 model using the training and validation sets to obtain the trained network model.

[0089] Specifically, the training environment configuration in this embodiment is as follows: operating system Windows 11, CPU AMD Ryzen 77840H, GPU NVIDIA GeForce RTX 4060 Laptop GPU, and software environment Python 3.10.18, PyTorch 2.0.0, and CUDA 11.8. Training parameter settings are: initial learning rate 0.01, optimizer stochastic gradient descent (SGD), batch size 16, training epochs 200, and input image resolution 720×720. Training process: The data-augmented training set is input into the LLDF-YOLO11 model, and the model performance is monitored in real time using the validation set. Network weights are adjusted, and the optimal model weights are saved after training.

[0090] Step 4: Use the trained LLDF-YOLO11 model to identify the bird's nest fruit flower image.

[0091] Specifically, the image of bird's nest, fruit, and flowers in a low-light environment to be detected is input into the trained LLDF-YOLO11 model. After one forward inference, the model outputs the bounding box coordinates and category of the target from the detection head, thus achieving accurate recognition of bird's nest, fruit, and flowers.

[0092] To verify the technical effectiveness of this invention, the test set of the constructed dataset was used as the test object to verify the object detection method of this invention. To comprehensively compare object detection performance, the results were presented from both visual and quantitative perspectives, using precision, recall, and mean average precision (mAP) for evaluation. Here, mAP refers to the average precision calculated by taking 10 IoU thresholds with a step size of 0.05 from 0.5 to 0.95. This invention uses mAP... 0.5-0.95 To represent, mAP is calculated as follows:

[0093]

[0094]

[0095]

[0096]

[0097] Where P is precision, R is recall; TP represents true positives, FP represents false positives, and FN represents false negatives; This represents the average precision function, where n represents the total number of categories.

[0098] In addition, one or both of the model parameter count (Params) and computational cost (FLOPs) are used as evaluation metrics for model lightweighting.

[0099] The constructed bird's nest fruit flower dataset has three target categories (flower buds, blooming flowers, and withered flowers). The dataset contains a total of 4260 images, with 3408 images in the training set, 426 in the validation set, and 426 in the test set. Figure 5 These are some typical images from the bird's nest fruit flower dataset. Figure 6 The detection images in the middle are comparison images of the target detection results obtained by the YOLOv11n algorithm and the present invention (red circles indicate missed detections, and red squares indicate false detections). Figure 6 The first to third columns represent the original image, the detection results of the YOLOv11n algorithm, and the detection results of this invention's embodiment, respectively. Figure 6 It can be seen that the embodiments of the present invention have stronger detection performance for bird's nest fruit flowers under low light and dense shading environment.

[0100] To verify the performance improvement effect of each improved module on the model, a series of ablation experiments were conducted on the test set under the same training environment and hyperparameter configuration. ICCFM, ADown, C2PSA_LSKA, and SCI_net were gradually introduced into the YOLOv11n baseline model for ablation experiments. The results of the relevant ablation experiments are shown in Table 1.

[0101] Table 1 Ablation Experiments of LLDF-YOLO11 Model

[0102]

[0103] Note: "√" indicates that this option is used; "-" indicates that this option is not used.

[0104] Table 1 shows the ablation experiments detailing the effectiveness of each improved module. First, the neck network of the baseline model YOLOv11n was replaced with ICCFM. Compared to the baseline model, this improvement reduced the number of model parameters and FLOPs by 30.2% and 15.6%, respectively. In terms of performance, the accuracy P remained unchanged, and mAP... 0.5-0.95 This represents a reduction of 1.4 percentage points, indicating that the ICCFM structure optimizes the model's lightweight nature.

[0105] Secondly, an ADown downsampling module was further introduced based on ICCFM. Compared with the baseline model, the combination of ICCFM and ADown reduced the number of parameters and FLOPs by 44.6% and 32.8%, respectively. Simultaneously, P and mAP... 0.5-0.95 These figures represent improvements of 0.4 and 2.3 percentage points, respectively. This indicates that the multi-scale feature fusion strategy of the ADown module effectively mitigates the loss of details caused by oversampling, significantly reducing computational costs while enhancing the model's detection accuracy.

[0106] Building upon this, the original C2PSA module was replaced with C2PSA_LSKA, aiming to capture broader contextual information with a lower parameter count. Compared to CCFM+ADown, the introduction of this module further fine-tuned the parameter count and FLOPs to 1.40M and 4.2G, respectively. In terms of accuracy, P and mAP... 0.5-0.95 Both increased by 0.4 percentage points.

[0107] Finally, SCI_net was introduced into layer 0 of the YOLO backbone network to calibrate low-light images of bird's nest fruit, improving the model's target recognition ability and forming the improved model LLDF-YOLO11. The introduction of SCI_net improved P and mAP. 0.5-0.95 The accuracy rates were further improved to 97.2% and 80.8% based on C2PSA_LSKA, representing increases of 1.1 and 1.7 percentage points respectively.

[0108] Compared to YOLOv11n, the final LLDF-YOLO11 model achieved improvements of 1.9 and 4.4 percentage points in P and mAP (0.5-0.95), respectively. In terms of lightweight design, FLOPs were reduced by 1.3G, and the number of parameters was reduced by 45.7%, fully validating the superiority of the improvement strategy. Specifically, the accuracy and efficiency metrics of the LLDF-YOLO11 model trained in this invention are as follows: precision reaches 97.2%, recall reaches 97.3%, and mean precision reaches 80.8%; the model has 1.4M parameters and 5.1G FLOPs.

[0109] Table 2 Comparison of LLDF-YOLO11 with other detection methods

[0110]

[0111] Test results show that the LLDF-YOLO11 model has 1.40M parameters, a 45.7% reduction compared to the baseline model YOLOv11n; accuracy is 97.2% and recall is 97.3%, representing improvements of 1.9 and 3.8 percentage points respectively compared to YOLOv11n; mAP 0.5-0.95 reaches 80.8%, an improvement of 4.4 percentage points compared to YOLOv11n; GFLOPs are only 5.2, lower than mainstream models such as Faster R-CNN, RT-DETR-L, and YOLOv8n.

[0112] Visualization results show that LLDF-YOLO11 has significantly lower false negative and false positive rates than the baseline model in low-light, occluded, and dense scenes.

[0113] The specific embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A lightweight identification method for bird's nest flowers in low-light environments based on LLDF-YOLO11, characterized in that, include: Step 1: Obtain the low-light image of the bird's nest fruit flower to be identified and the trained LLDF-YOLO11 model; the LLDF-YOLO11 model consists of three parts: the backbone network, the ICCFM neck network, and the detection head; Step 2: Input the low-light bird's nest and fruit flower image to be identified into the backbone network: The SCI_net module of the backbone network is used to perform illumination calibration to enhance the image quality and obtain the illumination-calibrated bird's nest and fruit flower image; The illumination-calibrated bird's nest and fruit flower image is passed through a hierarchical structure consisting of conv convolutional layers, multiple stacked "ADown downsampling module + C3k2 module", SPPF module, and C2PSA_LSKA module to extract multi-layer feature maps of the illumination-calibrated bird's nest and fruit flower image; Step 3: Input the multi-layer feature map output from the backbone network in Step 2 into the ICCFM neck network. The ICCFM neck network performs multi-scale fusion and enhancement on features from different levels to integrate contextual information and generate a multi-scale bird's nest, fruit and flower fusion feature map. Step 4: Input the multi-scale bird's nest fruit flower feature map obtained in Step 3 into the corresponding detection head to predict the position coordinates and category confidence of the bird's nest fruit flower in the map.

2. The lightweight identification method for bird's nest flowers in low-light environments based on LLDF-YOLO11 according to claim 1, characterized in that, The backbone network of the LLDF-YOLO11 model is specifically as follows: The SCI_net module is fused before layer 0 of the YOLOv11n basic framework; the standard convolutions in layers 1, 3, 5, and 7 of the traditional YOLOv11n backbone network are replaced with the ADown downsampling module. The ADown downsampling module includes an average pooling layer, channel splitting operation, a 3×3 convolutional layer, a max pooling layer, a 1×1 convolutional layer, and a concatenation layer; the input feature map is split into two paths after the average pooling layer extracts global information, and then splits into two paths after the channel splitting operation. One path is processed by a 3×3 convolutional layer to extract local details, and the other path is processed by a max pooling layer to capture salient features and then compressed by a 1×1 convolutional layer. Finally, the two feature paths are concatenated and output. The C2PSA module in the 10th layer of the traditional YOLOv11n backbone network is replaced with the C2PSA_LSKA module; the C2PSA_LSKA module combines the LSKA attention mechanism with the parallel spatial attention mechanism of the C2PSA module.

3. The lightweight identification method for bird's nest flowers in low-light environments based on LLDF-YOLO11 according to claim 1, characterized in that, The SCI_net module includes an illumination estimation module and a self-calibration module, which are connected in a cascaded manner. The training and testing processes of the SCI_net module are as follows: During the training of the SCI_net module, the output of the element-wise division operation between the original low-light bird's nest and fruit flower image and the current illumination image is input into the parameterization operator to generate a calibration term. The original low-light bird's nest and fruit flower image is then element-wise added to the calibration term to obtain the calibrated output, which serves as the input to the illumination estimation module. The input to the illumination estimation module is then processed by a convolution module to generate a residual term. The current illumination image is then element-wise added to the residual term to obtain the updated illumination image for the next stage. This iterative process continues until the next iteration. The process involves several stages, ultimately resulting in a trained illumination map. During the testing of the SCI_net module, the low-light bird's nest and fruit flower image to be tested is used as input to the illumination estimation module to generate an illumination map to be enhanced. The original low-light bird's nest and fruit flower image to be tested is divided element-wise with the illumination map to be enhanced to obtain the final enhanced output.

4. The lightweight identification method for bird's nest flowers in low-light environments based on LLDF-YOLO11 according to claim 1, characterized in that, The ADown downsampling module works as follows: the initial feature map is fed into the ADown downsampling module, where global information is first extracted and the dimensionality is reduced through an average pooling layer; then, the feature map is split into two branches along the channel dimension: one branch extracts local detail features through a 3×3 convolution; the other branch captures significant features through a max pooling layer, compresses the channels through a 1×1 convolution, and is given non-linear properties; finally, the features from the two branches are concatenated and fused along the channel dimension to obtain the downsampled feature map.

5. The lightweight identification method for bird's nest flowers in low-light environments based on LLDF-YOLO11 according to claim 1, characterized in that, The ICCFM neck network of the LLDF-YOLO11 model adopts an ICCFM structure, specifically: The multi-layer feature maps output by the backbone network are first preliminarily adjusted in terms of channel dimensions and features through convolutional layers to obtain dimension-matched feature maps to be fused. , , ; Future deep feature maps of autonomous backbone networks The resolution is increased by using the Upsample module, making it comparable to the feature maps of the layers in the autonomous backbone network. The two are spliced ​​and fused together, and then refined by the C3k2 module to generate an intermediate feature map; Subsequently, this intermediate feature map is further amplified in resolution by passing it through convolutional layers and an upsampling module, and then compared with feature maps from shallow layers of the backbone network. The splicing is completed along the channel dimension, and finally the features are integrated through the C3k2 module to generate a top-level small-scale feature map containing rich semantic and spatial information. The generated top-level small-scale feature map is input into the ADown module for downsampling, concatenated with the feature map of the intermediate feature map after convolution, and then fused by the C3k2 module to obtain the mid-level mesoscale feature map; next, this mid-level mesoscale feature map is fed into another ADown module for downsampling, and then combined with the generated... The images are stitched together and processed by the C3k2 module to obtain the underlying large-scale feature map of the large-scale receptive field.

6. The lightweight identification method for bird's nest flowers in low-light environments based on LLDF-YOLO11 according to claim 1, characterized in that, The training of the LLDF-YOLO11 model specifically involves: acquiring and filtering images of bird's nest fruit under low light conditions; enhancing and cropping the filtered bird's nest fruit images to construct a dataset; and training and validating the LLDF-YOLO11 model based on the training and validation sets in the dataset to obtain a well-trained LLDF-YOLO11 model with optimized accuracy and efficiency metrics.

7. The lightweight identification method for bird's nest flowers in low-light environments based on LLDF-YOLO11 according to claim 6, characterized in that, The collaborative optimization of accuracy-related indicators and efficiency-related indicators specifically involves: the accuracy-related indicator value reaching the preset threshold of the accuracy-related indicator, and the lightweight evaluation indicator value being lower than the lightweight evaluation indicator threshold.

8. The lightweight identification method for bird's nest flowers in low-light environments based on LLDF-YOLO11 according to claim 7, characterized in that, The accuracy-related indicator values ​​reaching the preset thresholds include: an accuracy value above the first preset threshold, a recall rate above the second preset threshold, and an average precision value above the third preset threshold; the lightweight evaluation indicator values ​​below the lightweight evaluation indicator thresholds include: the number of model parameters below the fourth preset threshold.

9. The lightweight identification method for bird's nest flowers in low-light environments based on LLDF-YOLO11 according to claim 6, characterized in that, The images of bird's nest fruit under low light conditions were collected during the following periods: 21:00-22:00 and 7:00-8:

00.

10. A lightweight identification system for bird's nests and fruit flowers in low-light environments based on LLDF-YOLO11, characterized in that, The module includes any one of the modules of the lightweight identification method for bird's nest and fruit flowers in low-light environments based on LLDF-YOLO11 as described in any one of claims 1-9.

Images