Chemical experiment device recognition method, device and equipment based on multi-layer feature fusion

Abstract

The application relates to a chemical experiment device recognition method, device and equipment based on multi-layer feature fusion, which comprises the following steps: pre-processing an acquired chemical experiment scene RGB image to obtain input features; adopting a backbone network to perform multi-scale feature extraction on the input features; performing dynamic channel calibration on the multi-scale features by adopting a channel attention module, respectively; performing deep enhancement and fusion on the channel-calibrated multi-scale features by adopting a multi-branch parallel processing at a neck to obtain multi-scale fusion features; and inputting the multi-scale fusion features into a decoder head to obtain a chemical experiment device recognition result. The method constructs a multi-layer feature fusion chemical experiment device recognition network, overcomes the challenges of multi-scale changes, component occlusion and inter-class similarity in chemical experiment device recognition in a complex experiment scene by developing advanced image recognition and target detection algorithms, and realizes real-time detection of chemical experiment devices in a real experiment scene.

Description

Technical Field

[0001] This invention belongs to the field of intelligent identification technology for chemical experimental devices, and relates to a method, device and equipment for identifying chemical experimental devices based on multi-layer feature fusion. Background Technology

[0002] In the process of achieving autonomous operation of chemical laboratory robots, the visual perception system, as its core perception center, is fundamental to enhancing its environmental interaction and task planning capabilities. Currently, general-purpose computer vision detection technology has made significant progress, and the emergence of large-scale pre-training technologies such as visual basic models has greatly enhanced the robustness and transferability of visual systems in diverse scenarios. However, in the highly structured and constrained professional setting of chemical experiments, robots still face significant cognitive bottlenecks in the refined identification, spatial state understanding, and operational intent reasoning of diverse experimental instruments, devices, and reagents, and a complete and reliable technical system has not yet been formed. To address the unique and complex nature of the chemical laboratory environment, developing intelligent robot systems with advanced environmental perception capabilities is crucial. Such systems need to integrate multimodal imaging and intelligent analysis technologies to achieve refined identification and dynamic understanding of experimental devices, material states, and reaction processes. Breakthroughs in this technology will not only replace researchers in performing highly repetitive and high-precision operational tasks, thereby improving the overall efficiency and reproducibility of experiments, but more profoundly, it will open up new possibilities for chemical scientific research and accelerate the paradigm shift in automated experiments.

[0003] Robot vision and imaging technologies have been widely applied in numerous industrial scenarios, including visual measurement, industrial quality inspection, target recognition, positioning and navigation, and automated assembly. However, the level of intelligence exhibited by existing systems, especially in comprehensive environmental perception and deep semantic cognition, still lags significantly behind human capabilities. This gap stems from a series of fundamental scientific challenges and key technological bottlenecks facing robot vision perception systems, the most prominent of which is insufficient robust understanding and adaptive decision-making capabilities for unstructured, dynamically changing, and complex scenarios. Summary of the Invention

[0004] To address the problems existing in the above-mentioned traditional methods, this invention proposes a method, apparatus, and device for identifying chemical experimental apparatus based on multi-layer feature fusion.

[0005] To achieve the above objectives, the embodiments of the present invention adopt the following technical solutions:

[0006] On the one hand, a method for identifying chemical experimental apparatus based on multi-layer feature fusion is provided, including the following steps:

[0007] Step 1: Acquire RGB images of the chemical experiment scene and preprocess them to obtain input features.

[0008] Step 2: Use a backbone network to extract multi-scale features from the input features to obtain multi-scale features.

[0009] Step 3: Perform dynamic channel calibration on the multi-scale features using the channel attention module to obtain the channel-calibrated multi-scale features.

[0010] Step 4: In the neck region, multi-branch parallel processing is used to perform deep enhancement and fusion of the multi-scale features after channel calibration to obtain multi-scale fused features. Specifically, the upper branch divides the multi-scale features after channel calibration into two paths: convolutional mapping and channel attention module calibration. After refining local features by the C2f module, the fusion is completed by the ADD module. The middle branch combines skip connections to complementarily fuse shallow detail features and deep semantic features after processing by convolution, channel attention module, and C2f module. The lower branch upsamples deep small-size features through a double deconvolution module to achieve size alignment of multi-scale features. The channel attention module is used to dynamically calibrate channel weights, suppress redundant background information, and highlight key feature channels of the chemical experimental device.

[0011] Step 5: Input the multi-scale fusion features into the decoder head to obtain the chemical experimental device recognition results.

[0012] On the other hand, a chemical experimental apparatus identification device based on multi-layer feature fusion is also provided, comprising:

[0013] The input feature determination unit is used to acquire RGB images of chemical experiment scenes and perform preprocessing to obtain input features.

[0014] The multi-scale feature extraction unit is used to extract multi-scale features from the input features using a backbone network to obtain multi-scale features.

[0015] The channel calibration unit is used to perform dynamic channel calibration on the multi-scale features using the channel attention module to obtain the channel-calibrated multi-scale features.

[0016] The multi-scale feature fusion unit is used to perform deep enhancement and fusion of channel-calibrated multi-scale features in the neck layer through multi-branch parallel processing to obtain multi-scale fused features. Specifically, the upper branch divides the channel-calibrated multi-scale features into two paths: convolutional mapping and channel attention module calibration. After refining local features by the C2f module, the fusion is completed by the ADD module. The middle branch combines skip connections to complementarily fuse shallow detail features and deep semantic features after processing by convolution, channel attention module, and C2f module. The lower branch upsamples deep small-size features through a double deconvolution module to achieve size alignment of multi-scale features. The channel attention module is used to dynamically calibrate channel weights, suppress redundant background information, and highlight key feature channels of the chemical experimental device.

[0017] The chemical experimental apparatus identification unit is used to input multi-scale fused features into the decoder head to obtain the chemical experimental apparatus identification result.

[0018] In another aspect, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the above-mentioned chemical experimental apparatus identification methods based on multi-layer feature fusion.

[0019] One of the above technical solutions has the following advantages and beneficial effects:

[0020] The aforementioned method, apparatus, and device for recognizing chemical experimental devices based on multi-layer feature fusion include: acquiring RGB images of a chemical experimental scene and preprocessing them to obtain input features; extracting multi-scale features from the input features using a backbone network to obtain multi-scale features; performing dynamic channel calibration on the multi-scale features using a channel attention module to obtain channel-calibrated multi-scale features; performing depth enhancement and fusion on the channel-calibrated multi-scale features through multi-branch parallel processing at the neck to obtain multi-scale fused features; and inputting the multi-scale fused features into the decoder head to obtain the chemical experimental device recognition result. This method constructs a multi-layer feature fusion network for recognizing chemical experimental devices. By developing advanced image recognition and target detection algorithms, it overcomes challenges such as multi-scale variations, component occlusion, and inter-class similarity in recognizing chemical experimental devices in complex experimental scenarios, achieving real-time detection of chemical experimental devices in real-world experimental scenarios. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of this application or the conventional technology, the drawings used in the description of the embodiments or the conventional technology will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart illustrating a chemical experimental apparatus identification method based on multi-layer feature fusion in one embodiment.

[0023] Figure 2 This is a schematic diagram of the structure of a chemical experimental apparatus identification model in one embodiment;

[0024] Figure 3 This is a schematic diagram of the channel attention module flow in one embodiment;

[0025] Figure 4 This is a schematic diagram illustrating the working principle of the deconvolution module in one embodiment, wherein... Figure 4 Image (a) is a schematic diagram of a normal convolution operation. Figure 4 (b) is a schematic diagram of the deconvolution operation;

[0026] Figure 5 Here is a training curve plot on the ChemEq25 dataset in one embodiment, where Figure 5 (a) to (c) represent the training set of the ChemEq25 dataset, respectively. , , A schematic diagram of the loss curve. Figure 5 (d) through (f) represent the validation sets of the ChemEq25 dataset, respectively. , , A schematic diagram of the loss curve. Figure 5 Figures (g) to (j) show the performance metrics of the model on the validation set of the ChemEq25 dataset: precision, recall, MAP50, and Map50-95, respectively.

[0027] Figure 6 This is a training plot on a dataset of labeled chemical apparatus images in one embodiment, wherein... Figure 6 (a) to (c) are the training set of the labeled chemical apparatus image dataset, respectively. , , A schematic diagram of the loss curve. Figure 6 (d) through (f) represent the validation set of the labeled chemical apparatus image dataset, respectively. , , A schematic diagram of the loss curve. Figure 6 Figures (g) to (j) show the performance metrics of the model on the validation set of the labeled chemical device image dataset, namely precision, recall, MAP50, and Map50-95.

[0028] Figure 7This is a schematic diagram of the visualization results on a dataset of labeled chemical apparatus images in one embodiment;

[0029] Figure 8 This is a schematic diagram of the visualization results on the ChemEq25 dataset in one embodiment. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0031] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the application.

[0032] It should be noted that reference to embodiments herein means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The presentation of this phrase in various locations throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. Those skilled in the art will understand that the embodiments described herein can be combined with other embodiments. Terms used herein refer to and / or mean any combination of one or more of the associated listed items, and all possible combinations thereof.

[0033] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0034] In one embodiment, such as Figure 1 As shown, a method for identifying chemical experimental apparatus based on multi-layer feature fusion is provided, which may include the following processing steps 1 to 5:

[0035] Step 1: Acquire RGB images of the chemical experiment scene and preprocess them to obtain input features.

[0036] Specifically, the model takes a 640×640×3 RGB image of a chemical experiment scene as input.

[0037] Step 2: Use a backbone network to extract multi-scale features from the input features to obtain multi-scale features.

[0038] Specifically, the backbone network achieves multi-stage feature extraction through stacked convolution (Conv), lightweight residual module C2f, and spatial pyramid pooling (SPPF): the C2f module enhances gradient propagation while maintaining computational efficiency through branched residual structures, effectively capturing the local texture and contour features of glass instruments; the SPPF module aggregates multi-scale information from micro-scale scales to macro-scale instrument structures through multi-scale pooling and feature concatenation.

[0039] Step 3: Perform dynamic channel calibration on the multi-scale features using the channel attention module to obtain the channel-calibrated multi-scale features.

[0040] Specifically, a channel attention (CA) module is embedded after features at different scales. By explicitly modeling the dependencies between channels, the feature channel weights are adaptively calibrated to highlight feature channels that are strongly related to the chemical experimental setup (such as high-frequency channels corresponding to the light transmittance of glass materials) and suppress interference from redundant information such as the experimental platform and background markings.

[0041] Step 4: In the neck region, multi-branch parallel processing is used to perform deep enhancement and fusion of the multi-scale features after channel calibration to obtain multi-scale fused features. Specifically, the upper branch divides the multi-scale features after channel calibration into two paths: convolutional mapping and channel attention module calibration. After refining local features by the C2f module, the fusion is completed by the ADD module. The middle branch combines skip connections to complementarily fuse shallow detail features and deep semantic features after processing by convolution, channel attention module, and C2f module. The lower branch upsamples deep small-size features through a double deconvolution module to achieve size alignment of multi-scale features. The channel attention module is used to dynamically calibrate channel weights, suppress redundant background information, and highlight key feature channels of the chemical experimental device.

[0042] Specifically, the neck, as the core connecting unit between the backbone network and the decoder head, first receives multi-scale feature maps from the backbone. Through multi-branch parallel processing, it achieves in-depth feature enhancement and fusion: the upper branch divides the input features into two paths: Conv mapping and CA channel calibration. After C2f refinement of local features, fusion is completed via ADD; the middle branch, combined with skip connections, complementaryly fuses shallow detail features and deep semantic features after Conv, CA, and C2f processing; the lower branch upsamples deep, small-sized features through double CT deconvolution to achieve size alignment of multi-scale features; the CA module embedded throughout dynamically calibrates channel weights, suppressing redundant background information and highlighting key feature channels of the chemical experimental apparatus; the C2f module preserves gradient propagation through a lightweight residual structure, accurately capturing instrument contours and texture details; finally, through multiple ADD operations... Element-wise fusion and cross-layer jump connections enable the neck module to output high-quality fusion features that combine microscopic details (such as instrument scales and tube outlines) with macroscopic semantics (such as instrument type and overall structure). This provides strong discriminative feature support for the identification of chemical experimental devices in the subsequent decoding stage, effectively improving the robustness and accuracy of identification in complex experimental scenarios (such as multiple instrument stacks and uneven lighting).

[0043] By employing multi-scale feature fusion, the entire process integrates deep and shallow features, ensuring that feature information from each level is mutually integrated, thereby enabling shallow features to possess powerful semantic information.

[0044] By constructing a Channel Attention (CA) module and a Deconvolution (CT) module, high-level semantic features are used as weights for low-level feature filtering. These filtered features are then added point-by-point to the high-level semantic features to achieve multi-scale feature fusion, thereby ultimately improving the model's feature representation capability. Compared to fixed, parameter-free upsampling methods such as bilinear interpolation and nearest neighbor interpolation, the learnable weights of deconvolution make it a learnable upsampling mechanism. During training, the network can automatically learn how to optimally 'unfold' deep semantic information onto a larger spatial scale, which typically generates higher-quality, more semantically coherent, and more beneficial upsampled feature maps for subsequent tasks, thus improving the fusion effect.

[0045] Step 5: Input the multi-scale fusion features into the decoder head to obtain the chemical experimental device recognition results.

[0046] Specifically, this decoder head, as the final prediction unit of the chemical experimental apparatus recognition model, receives the multi-scale fusion features output from the neck. Its core task is to map the abstract feature space into pixel-level or detection-level results that can be directly used for instrument recognition. This ensures that the fine contours of the chemical experimental apparatus (such as burette pistons and pipette scales) and small components (such as droppers and tubing interfaces) are not obscured. The decoder head also embeds convolution and lightweight feature refinement modules to further calibrate feature channel weights and enhance the distinction between instrument semantic information and the background. Finally, through detection branches (classification + regression) or segmentation branches, it outputs the category label, precise bounding box coordinates, and even instance masks of the chemical experimental apparatus. This enables high-precision recognition and localization of chemical experimental apparatus in complex experimental scenarios (multiple instrument stacks, uneven lighting, component occlusion), providing direct decision-making basis for intelligent laboratory monitoring and compliance testing of experimental operations.

[0047] The decoder head integrates a dynamic query mechanism for fine-grained decoding and interaction of candidate region features, thereby further improving classification and localization accuracy while maintaining the overall efficiency of the model.

[0048] The decoder head uses deformable self-attention, which significantly reduces computational cost. This part also utilizes CUDA operators to accelerate execution. The decoder, with an auxiliary prediction head, iteratively optimizes object queries to generate bounding boxes and confidence scores.

[0049] In the chemical experimental apparatus recognition method based on multi-layer feature fusion, the decoder head acts as the final prediction unit. Its main function is to transform the multi-scale fused features output from the neck into specific results usable for chemical experimental apparatus recognition, such as category labels and bounding box coordinates. Its operation process includes receiving multi-scale fused features, performing efficient computation using deformable self-attention and CUDA operators, refining and interacting with candidate region features through a dynamic query mechanism to improve classification and localization accuracy, and finally outputting high-precision chemical experimental apparatus recognition results through detection or segmentation branches.

[0050] Steps 2 to 5 constitute a chemical experimental apparatus identification model based on multi-layer feature fusion. The structure of the chemical experimental apparatus identification model is as follows: Figure 2 As shown.

[0051] The aforementioned method for recognizing chemical experimental apparatus based on multi-layer feature fusion includes: acquiring RGB images of a chemical experimental scene and preprocessing them to obtain input features; using a backbone network to extract multi-scale features from the input features to obtain multi-scale features; applying a channel attention module to dynamically calibrate the multi-scale features to obtain channel-calibrated multi-scale features; performing depth enhancement and fusion on the channel-calibrated multi-scale features through multi-branch parallel processing at the neck to obtain multi-scale fused features; and inputting the multi-scale fused features into the decoder head to obtain the chemical experimental apparatus recognition result. This method constructs a multi-layer feature fusion network for recognizing chemical experimental apparatus. By developing advanced image recognition and target detection algorithms, it overcomes challenges such as multi-scale variations, component occlusion, and inter-class similarity in complex experimental scenarios, achieving real-time detection of chemical experimental apparatus in real-world experimental settings.

[0052] In one embodiment, the backbone network in step 2 is the backbone network of the YOLOv8 network model.

[0053] In one embodiment, such as Figure 3 As shown, the channel attention module includes a global max pooling layer, a global average pooling layer, and a sigmoid function. The specific process of using the channel attention module for dynamic channel calibration in step 3 includes: processing the features of the input channel attention module through the global max pooling layer and the global average pooling layer respectively, then adding and fusing them; activating the fused result with the sigmoid function to obtain the channel attention features; and processing the channel attention features and the features of the input channel attention module through a channel-by-channel multiplication operation to obtain the channel-calibrated features.

[0054] Specifically, the Channel Attention (CA) module enhances the network's ability to extract key information by adaptively calibrating the importance of feature channels. Its concise design plays a crucial role as an intelligent feature filter in chemical experimental apparatus recognition tasks. Its core function is to guide the model to adaptively determine and strengthen the color channels in the image that are most effective for recognizing chemical experimental apparatus, thereby significantly improving the model's discriminative ability. Its processing flow begins with the input feature map, which is fed in parallel into two branches: global max pooling and global average pooling. Max pooling focuses on extracting the most salient activation response within each channel to capture the most discriminative local features, acting like a feature equalizer that focuses on capturing the most salient and strongest feature response within each channel. This helps identify instrument highlights, special textures, or unusual colors, such as reflections at glass edges or the color of specific reagents. Average pooling, on the other hand, aggregates the overall statistical information of the channels to preserve the global contextual distribution of features, acting like a feature equalizer, calculating the overall average value for each channel to obtain global, smooth background and material information. This helps in understanding the overall tone, material (such as the transparency of glass), and stable ambient lighting of the instrument. Both pooling operations provide complementary perspectives on feature representation while achieving dimensionality reduction and compressing redundant information.

[0055] Subsequently, the two channel description vectors obtained through pooling are fused and input into the Sigmoid activation function to generate a channel weight vector with values ​​between 0 and 1, as shown in the following two formulas:

[0056] F∈R C×H×W

[0057] FCA∈R C×1×1

[0058] Where F is the input, FCA∈R C×1×1 This represents the features after channel attention. The weight vector quantifies the importance of each channel to the current task. Finally, this weight vector is recalibrated with the original input feature map through channel-wise multiplication, thereby enhancing the response of important channels and suppressing interference from less important channels. By combining the advantages of max pooling and average pooling, this module can adaptively select and strengthen information-rich channels in the feature map with low computational overhead, thus improving the discriminative ability of subsequent processing.

[0059] In one embodiment, the neck region includes: seven convolutional modules, six channel attention modules, two deconvolutional modules, four C2f modules, and four ADD modules; step 4 includes: processing the channel-calibrated small-sized features through the first convolutional module to obtain a first convolutional feature; processing the first convolutional feature through the first channel attention module for dynamic channel calibration, then processing it through the second convolutional module to obtain a second convolutional feature; upsampling the first convolutional feature through the first deconvolutional module to obtain a first upsampled feature; inputting the first upsampled feature into the second channel attention module, multiplying the obtained channel attention feature with the channel-calibrated mesoscale feature mapped by the third convolutional module, and inputting the result with the first upsampled feature into the first ADD module to obtain a first fused feature; processing the first fused feature through the first C2f module, then performing dynamic channel calibration using the third channel attention module, and then processing it through the fourth convolutional module to obtain a third convolutional feature; upsampling the first upsampled feature using the second deconvolutional module to obtain a second upsampled feature; and so on. The upsampled features are input into the fourth channel attention module. The resulting channel attention features are multiplied by the channel-calibrated large-size features processed by the fifth convolution module. The result and the second upsampled feature are input into the second ADD module to obtain the second fused feature. The second fused feature is input into the second C2f module to obtain the large-scale fused feature. The large-scale fused feature is input into the sixth convolution module to obtain the fourth convolution feature. The fourth convolution feature is input into the fifth channel attention module. The resulting attention feature is multiplied by the third convolution feature. The result and the fourth convolution feature are input into the third ADD module. The result is input into the third C2f module to obtain the medium-scale fused feature. The fourth convolution feature is input into the seventh convolution module to obtain the fifth convolution feature. The fifth convolution feature is input into the sixth channel attention module. The resulting channel attention feature is multiplied by the second convolution feature. The result and the fifth convolution feature are input into the fourth ADD module. The result is input into the fourth C2f module to obtain the small-scale fused feature.

[0060] Specifically, deconvolution (also known as transposed convolution) is the core learnable operation for upsampling feature maps, and its process can be likened to intelligent feature weaving. For example... Figure 4 As shown, Figure 4 (a) defines the basic elements of the operation. The square base at the bottom symbolizes the low-resolution feature map of the input, and the image of the chemical experimental apparatus on it indicates that the input contains semantic information. The blue square plate above represents the convolution kernel, which is the core computational unit that can be learned. The rope connecting the two symbolically indicates that the convolution kernel will act on the input to perform corresponding calculations. Figure 4(b) Dynamically illustrates the computational process. The small blue squares below represent the effective feature points in a larger spatial grid after zero-value padding of the input features. The gridded platform itself represents the sparse grid formed after the system inserts zero values ​​to expand the size. The black plate above and its blurred chemical image are the final high-resolution feature map output. This operation is crucial in chemical experimental apparatus recognition tasks. Because convolutional neural networks lose spatial details at deeper layers, the features of tiny or finely structured instruments (such as burette scales or beaker rims) are easily blurred. Deconvolution, through the inverse convolution path, adaptively and intelligently reconstructs the semantic information of what the instrument is—understood by the deep network—into the precise outline of the instrument and the spatial details of where its components are located, thus achieving a mapping from abstract concepts to concrete locations. It not only recovers the details lost due to downsampling but also, when fused with shallow high-resolution features, provides the model with the ability to simultaneously understand the semantics and precise geometry of the instrument, thereby significantly improving the recognition and positioning accuracy of chemical experimental apparatuses at various scales.

[0061] In one embodiment, the convolution module includes a convolutional layer, a batch normalized layer, and a SiLU activation function.

[0062] In one embodiment, the decoder head is used to further calibrate the feature channel weights using convolutional layers and a lightweight feature refinement module, enhancing the distinction between instrument semantic information and background, and outputting the category label and precise bounding box coordinates of the chemical experimental apparatus by detecting or segmenting branches.

[0063] In a validation embodiment, the effectiveness of the method was validated on the ChemEq25 dataset and a dataset of labeled chemical apparatus images.

[0064] The ChemEq25 dataset contains 4,599 images of chemical experimental equipment, all of which have been professionally labeled. Aimed at supporting machine learning research for real-time equipment detection, this dataset includes realistic experimental scenarios captured from multiple angles, under various lighting conditions, and against complex backgrounds, helping models achieve stronger generalization and stability in practical deployments. Furthermore, all images have undergone size normalization and format standardization preprocessing and have been divided into training, validation, and test sets, providing high-quality, reproducible benchmark data for chemical equipment identification and detection tasks.

[0065] This labeled chemical apparatus image dataset contains 5,078 image samples processed using a standardized annotation process. All images were extracted from experimental video frames recorded by smartphones. This dataset specifically supports object detection tasks, with annotations covering the experimenter's hand area in the experimental scene, as well as the location coordinates and category labels of six common chemical apparatus types: conical beakers, eggplant flasks, Erlenmeyer flasks, pipettes, reagent bottles, and separatory funnels. All images in the dataset have been uniformly converted to JPG format, including three resolutions: 1920×1080, 1280×720, and 960×540. Training, validation, and test sets have been completed, providing standardized data support for the algorithm development and performance evaluation of related detection models.

[0066] L1 loss (also known as Mean Absolute Error, MAE) is widely used in regression tasks. This loss function measures error by calculating the sum of the absolute differences between the model's predicted values ​​and the true values ​​(as shown in the formula). Compared to L2 loss (Mean Squared Error, MSE), L1 loss is more robust to outliers because it is insensitive to extreme values ​​in the data. This characteristic makes L1 loss typically provide more robust performance when dealing with datasets containing outliers.

[0067] The L1 loss function, by calculating the absolute difference (rather than the squared difference) between the predicted and true values, exhibits low sensitivity to outliers. This characteristic makes L1 loss more robust to extreme values ​​in the dataset. Furthermore, the gradient of L1 loss remains constant at ±1, which can lead some optimization methods to frequently update parameters, resulting in sparse adjustments. However, L1 loss effectively reflects the average magnitude of the prediction error and maintains the same units of measurement as the original data, significantly improving interpretability. In scenarios such as image denoising, L1 loss helps maintain the accuracy of pixel values ​​in an image, effectively avoiding imposing excessive penalties on large errors.

[0068] Generalized Intersection over Union (GIoU) ​​loss is an advanced loss function widely used in object detection tasks. Compared to standard GIoU loss, GIoU loss demonstrates superior performance in optimizing bounding box localization accuracy, especially when bounding boxes do not overlap, consistently providing effective optimization signals. Its application in object detection models significantly improves model performance in bounding box regression tasks. Within this framework, the C value represents the area of ​​the minimum bounding rectangle, which must contain both the predicted and ground truth bounding boxes; this calculation is called the closure area. The expression for GIoU loss is:

[0069] ;

[0070] ;

[0071] ;

[0072] in, For intersection, union, and comparison, For generalized intersection and comparison, For generalized intersection and comparison of losses.

[0073] Generalized intersection and union ratio loss ( The GIoU value ranges from -1 to 1. When two bounding boxes completely overlap, the GIoU value is 1. If the bounding boxes do not overlap, the GIoU value drops below 0. Therefore, the GIoU loss value reflects the higher the alignment between the predicted and ground truth bounding boxes; the lower the value. Unlike the standard Cross-Union Ratio (CIRR) loss, the GIoU loss provides an optimization signal even when the bounding boxes do not overlap, thus accelerating model convergence. Compared to CIRR loss, the GIoU loss provides more stable gradient updates, which helps to more accurately regress the target location.

[0074] During the training of an object detection model, the classification loss ( The classification loss plays a crucial role in ensuring the model can correctly identify objects. It's an intelligent learning strategy that identifies which sample targets the model can already accurately judge, reducing focus on those that the model cannot yet distinguish, and concentrating on optimizing other samples that are critical to improving overall performance. This directly enables the model to maintain high average accuracy while also handling various complex scenarios to a certain extent. It drives the classifier to learn the essential features of various targets by quantifying the difference between the model's predicted class probability distribution and the true labels, addressing the two inherent challenges of extremely imbalanced background complexity and significant differences between easy and difficult samples in object detection.

[0075] The collaborative optimization of classification loss, generalized intersection-union-ratio loss, and L1 loss constitutes the multi-task learning framework of the object detection model. This ensures that while identifying the category, the model can accurately regress the target location and assess the existence confidence, ultimately achieving reliable end-to-end detection performance.

[0076] This embodiment comprehensively evaluates the model's efficiency using different benchmark datasets. To enhance the transparency of the evaluation method, these datasets and experimental conditions are described in detail, verifying the model's effectiveness on the test datasets. During the evaluation, key metrics such as recognition accuracy, recall, mAP@0.5, and mAP@0.5:0.95 were selected, as these metrics effectively reflect the model's detection capabilities. Furthermore, GFLOPs (gigaflops) and FPS (frames per second) were used as evaluation metrics to measure the model's computational complexity.

[0077] This method demonstrates superior and balanced detection performance on the ChemEq25 dataset, with its advantages manifested in three aspects. First, the model achieves a very high level of practicality under conventional evaluation criteria, with mAP50 exceeding 0.95 for the vast majority of categories, and even reaching above 0.99 for some categories. Specifically, the mAP50 for categories such as precision electronic balances, porcelain mortars and pestles, and burette racks reaches as high as 0.995, indicating that the model achieves near-error-free identification of instruments with clear structures. Second, the model exhibits strong feature generalization and detail discrimination capabilities, particularly in the accurate differentiation of highly similar variant instruments. For example, the mAP50 for single-necked, two-necked, and three-necked round-bottom flasks reaches 0.974, 0.984, and 0.985, respectively, proving that the model does not merely learn basic outlines but accurately captures the key functional feature of the number of interfaces. Finally, the model achieved an excellent balance between precision and recall, with both metrics exceeding 0.95 for most categories. This means the model can report targets with high confidence and effectively cover the vast majority of instruments to be inspected in the scenario, significantly reducing the risk of false alarms and missed detections in practical applications. In summary, these results define a new benchmark for chemical experimental apparatus recognition from three dimensions: accurate recognition, clear differentiation, and complete identification, providing crucial technical validation for building truly reliable and usable laboratory automation vision systems. Experimental results on the ChemEq25 dataset are shown in Table 1.

[0078] Table 1. Experimental results using the ChemEq25 dataset.

[0079]

[0080] Figure 5 The training curves shown systematically reveal the high stability and excellent performance of the model during convergence on the ChemEq25 dataset. Figure 5 (a) to (c) represent the training set of the ChemEq25 dataset, respectively. , , A schematic diagram of the loss curve. Figure 5 (d) through (f) represent the validation sets of the ChemEq25 dataset, respectively. , , A schematic diagram of the loss curve. Figure 5 Figures (g) to (j) show the performance metrics of the model on the validation set of the ChemEq25 dataset: precision (B), recall (B), MAP50 (B), and Map50-95 (B), respectively. All core loss curves ( , , The model exhibits a smooth, synchronous, monotonically decreasing trend on both the training and validation sets, eventually stabilizing at a low value plateau. This indicates that the optimization process is highly efficient and robust, without gradient anomalies or overfitting. Meanwhile, the key validation set performance metrics depict a clear path of continuous model enhancement: precision and recall steadily rise to high levels (approaching 0.99 and 0.98 respectively), demonstrating that the model improves both discrimination accuracy and target coverage. Importantly, the combined performance metrics mAP50 and mAP50-95 also show a consistent and significant upward trend, ultimately converging to excellent levels. This directly confirms the systematic optimization of the model's comprehensive detection capabilities in localization and classification. Overall, these smooth, non-oscillating curves, all pointing towards the optimal solution, constitute a textbook example of successful training. From a dynamic process perspective, they irrefutably verify that the model learned highly generalized and discriminative feature representations on the ChemEq25 dataset, providing solid and credible procedural evidence for its top-tier static performance shown in the table.

[0081] The model demonstrated exceptional environmental adaptability and practical robustness on a more complex and realistic dataset of labeled chemical apparatus images, with its superiority particularly evident in its accurate resolution of open and dynamic experimental scenarios. Although the overall metrics declined compared to the structurally standardized ChemEq25 dataset, this reflects the real-world challenges of that dataset—including operator hand interactions, more complex backgrounds, and multi-object occlusion. Under these conditions, the model still achieved an extremely high accuracy of 0.961 for the separatory funnel, indicating its reliable recognition of structurally complex devices. Simultaneously, the recall rate (0.838) and mAP50 (0.825) for reagent bottles were outstanding, demonstrating the model's highly sensitive perception of target presence in cluttered backgrounds. Crucially, the model successfully distinguished between hands and experimental apparatus (achieving an mAP50 of 0.78 for the hand category), achieving a fundamental understanding of human-object interaction scenarios and representing a significant step towards practical deployment. All indicators are balanced and significantly higher than the random baseline, indicating that the model is not limited to processing specimen-like images under ideal conditions, but has the strong potential for reliable visual perception and analysis in real, open chemical experimental environments. The experimental results for the labeled chemical apparatus image dataset are shown in Table 2.

[0082] Table 2 Experimental Results of the Annotated Chemical Apparatus Image Dataset

[0083]

[0084] Figure 6 The training curves shown powerfully demonstrate, from a dynamic process perspective, that the model exhibits excellent learning stability, efficient convergence, and strong generalization performance on a highly complex dataset of labeled chemical apparatus images, which is its core superiority. Figure 6 (a) to (c) are the training set of the labeled chemical apparatus image dataset, respectively. , , A schematic diagram of the loss curve. Figure 6 (d) through (f) represent the validation set of the labeled chemical apparatus image dataset, respectively. , , A schematic diagram of the loss curve. Figure 6 Figures (g) to (j) show the performance metrics of the model on the validation set of a labeled chemical apparatus image dataset: precision, recall, MAP50, and Map50-95, respectively. Specifically, all loss curves ( , , The training and validation sets both exhibit a smooth, synchronous, monotonically decreasing trend, eventually converging tightly to a low level. This classic, non-divergent double-line convergence clearly indicates that the model did not overfit during the optimization process, and the learned features possess extremely strong generalization ability. Simultaneously, the precision, recall, and mAP metrics, reflecting the final task performance, steadily and collaboratively increased from the early stages of training, eventually stabilizing at a high level. This intuitively demonstrates that the model's comprehensive capabilities in target discovery, discrimination, and localization have been systematically and reliably constructed when dealing with human-hand interactions, complex occlusions, and cluttered backgrounds in real-world scenarios. In summary, these near-textbook-perfect ideal training curves not only validate the effectiveness of the algorithm and training strategy in the process but also dynamically reveal the model's ability to robustly extract robust feature representations from highly challenging real-world chemical experiment images, thus providing a solid and credible internal basis for its excellent static detection results on this dataset.

[0085] Table 3 shows the experimental results, systematically demonstrating the comprehensive superiority of this method on two chemical experimental apparatus datasets. Its outstanding achievement lies in achieving near-perfect performance in standardized scenarios while maintaining a reliable and efficient baseline capability in complex real-world scenarios, all while maintaining a lightweight design. Specifically, on the highly standardized ChemEq25 dataset, the model achieved breakthrough detection accuracy (97.90% mAP50) and astonishing processing speed (204.1 FPS), achieving a perfect balance between accuracy and speed. This proves the ultimate effectiveness of the model architecture and training strategy under ideal conditions. Crucially, on a more realistic dataset of labeled chemical apparatus images with complex interference, the model still maintains highly practical performance (72.91% mAP50, 196.1 FPS), with its parameter count only slightly decreasing to 5.26M. This performance confirms that the model is not a greenhouse model; its learned features possess strong generalization ability and scene adaptability, providing a high-performance, highly available unified visual perception foundation for different automation needs, from structured laboratories to open workbenches.

[0086] Table 3. Training results using the ChemEq25 dataset and the labeled chemical apparatus image dataset.

[0087]

[0088] In the visualization experiment of annotated chemical apparatus image dataset, the visualization results systematically demonstrate the superior perception, analysis, and adaptation capabilities of the proposed model in real and complex chemical experimental scenarios. Its superiority is reflected in three progressive levels. First, the model possesses near-perfect recognition accuracy for standard chemical experimental apparatus, such as... Figure 7 The model achieves a confidence level of 1.0 for round-bottom flasks, demonstrating its reliability in standard scenarios. Secondly, the model exhibits strong robustness in highly complex, unstructured environments, such as... Figure 7 As shown, it can accurately segment and focus key devices on a real experimental platform with cluttered instruments and strong background interference, concentrating its attention (in the form of heatmaps) on the most discriminative areas (such as the graduations on a graduated cylinder and the liquid level). Crucially, the model has initially achieved semantic understanding of dynamic human-object interaction scenarios. It can not only recognize the operator's hands but also annotate key parts interacting with containers using heatmaps. This signifies that the model's perception has moved from static object recognition to understanding experimental procedures. In summary, these visualization results collectively confirm that this vision system has transcended simple instrument recognition, evolving into an intelligent experimental eye capable of understanding complex laboratory scenarios, adapting to dynamic operations, and providing highly interpretable perceptual results, providing a vital foundation for contextual awareness in chemical automation.

[0089] The visualization results of the ChemEq25 dataset are as follows: Figure 8 As shown, the visualization results of this ChemEq25 dataset collectively demonstrate the superior performance of the proposed model from multiple dimensions. First, as... Figure 8 As shown, the model achieves a confidence level of 0.9 for recognizing typical instruments such as beakers and conical flasks in standard scenarios, with accurate bounding box localization. This intuitively demonstrates that it possesses near-perfect detection accuracy and reliability on high-resolution, structurally sound images, laying a solid foundation for automated applications. More importantly, Figure 8 The attention heatmaps revealed by the demonstration showcase the model's deep understanding capabilities beyond surface-level recognition: instead of processing images equally, it intelligently focuses on the most discriminative key regions (such as specific reaction areas and instrument feature structures). This attention mechanism effectively filters out interference in complex backgrounds. Particularly noteworthy is that even in cluttered real-world experimental scenarios, the model's thermal response clearly and coherently delineates the core target, clearly distinguishing it from irrelevant background objects (such as coffee machines and laptops). This strongly demonstrates the system's robust anti-interference capabilities and scene generalization robustness. In summary, these visualizations not only demonstrate high-precision static recognition results but also dynamically reveal the model's intelligent, focused, and robust perceptual process, indicating that it has evolved from a model that works well on ideal datasets into an intelligent chemical eye capable of understanding complex real-world scenes and possessing high practical value.

[0090] It should be understood that, although the above Figure 1 The steps are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise explicitly stated in this document, there is no strict order in which these steps are executed; they can be performed in other orders. Furthermore, the above... Figure 1 At least some of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0091] In one embodiment, a chemical experimental apparatus identification device based on multi-layer feature fusion is also provided, comprising:

[0092] The input feature determination unit is used to acquire RGB images of chemical experiment scenes and perform preprocessing to obtain input features.

[0093] The multi-scale feature extraction unit is used to extract multi-scale features from the input features using a backbone network to obtain multi-scale features.

[0094] The channel calibration unit is used to perform dynamic channel calibration on the multi-scale features using the channel attention module to obtain the channel-calibrated multi-scale features.

[0095] The multi-scale feature fusion unit is used to perform deep enhancement and fusion of channel-calibrated multi-scale features in the neck layer through multi-branch parallel processing to obtain multi-scale fused features. Specifically, the upper branch divides the channel-calibrated multi-scale features into two paths: convolutional mapping and channel attention module calibration. After refining local features by the C2f module, the fusion is completed by the ADD module. The middle branch combines skip connections to complementarily fuse shallow detail features and deep semantic features after processing by convolution, channel attention module, and C2f module. The lower branch upsamples deep small-size features through a double deconvolution module to achieve size alignment of multi-scale features. The channel attention module is used to dynamically calibrate channel weights, suppress redundant background information, and highlight key feature channels of the chemical experimental device.

[0096] The chemical experimental apparatus identification unit is used to input multi-scale fused features into the decoder head to obtain the chemical experimental apparatus identification result.

[0097] In one embodiment, the backbone network in the multi-scale feature extraction unit is the backbone network of the YOLOv8 network model.

[0098] In one embodiment, the channel attention module includes a global max pooling layer, a global average pooling layer, and a sigmoid function. The specific process of using the channel attention module for dynamic channel calibration in the channel calibration unit includes: processing the features of the input channel attention module through the global max pooling layer and the global average pooling layer respectively, then adding and fusing them; activating the fused result with the sigmoid function to obtain the channel attention features; and processing the channel attention features and the features of the input channel attention module through a channel-by-channel multiplication operation to obtain the channel-calibrated features.

[0099] In one embodiment, the neck region includes: seven convolutional modules, six channel attention modules, two deconvolutional modules, four C2f modules, and four ADD modules; a multi-scale feature fusion unit is further configured to: process the channel-calibrated small-sized features through the first convolutional module to obtain a first convolutional feature; process the first convolutional feature through the first channel attention module for dynamic channel calibration, then through the second convolutional module to obtain a second convolutional feature; upsample the first convolutional feature through the first deconvolutional module to obtain a first upsampled feature; input the first upsampled feature into the second channel attention module, multiply the obtained channel attention feature with the channel-calibrated mid-scale features mapped by the third convolutional module, and input the result and the first upsampled feature into the first ADD module to obtain a first fused feature; process the first fused feature through the first C2f module, then perform dynamic channel calibration using the third channel attention module, and then process it using the fourth convolutional module to obtain a third convolutional feature; and upsample the first upsampled feature using the second deconvolutional module to obtain a second upsampled feature. The second upsampled feature is input into the fourth channel attention module. The resulting channel attention feature is multiplied by the channel-calibrated large-size feature processed by the fifth convolution module. The result and the second upsampled feature are then input into the second ADD module to obtain the second fusion feature. The second fusion feature is input into the second C2f module to obtain the large-scale fusion feature. The large-scale fusion feature is input into the sixth convolution module to obtain the fourth convolution feature. The fourth convolution feature is input into the fifth channel attention module. The resulting attention feature is multiplied by the third convolution feature. The result and the fourth convolution feature are input into the third ADD module. The result is then input into the third C2f module to obtain the medium-scale fusion feature. The fourth convolution feature is input into the seventh convolution module to obtain the fifth convolution feature. The fifth convolution feature is input into the sixth channel attention module. The resulting channel attention feature is multiplied by the second convolution feature. The result and the fifth convolution feature are then input into the fourth ADD module. The result is then input into the fourth C2f module to obtain the small-scale fusion feature.

[0100] In one embodiment, the convolution module includes a convolutional layer, a batch normalized layer, and a SiLU activation function.

[0101] In one embodiment, the decoder head in the chemical experimental apparatus identification unit is used to further calibrate the feature channel weights using convolutional layers and a lightweight feature refinement module, thereby enhancing the distinction between the instrument's semantic information and the background. By detecting branches or segmenting branches, it outputs the category label and precise bounding box coordinates of the chemical experimental apparatus.

[0102] It is understood that for a detailed explanation of the chemical experimental apparatus identification device based on multi-layer feature fusion, please refer to the corresponding explanations of the various embodiments of the chemical experimental apparatus identification method based on multi-layer feature fusion described above, and will not be repeated here. Each module in the aforementioned chemical experimental apparatus identification device based on multi-layer feature fusion can be implemented entirely or partially through software, hardware, or a combination thereof. Each module can be embedded in hardware or independently of a device with data processing capabilities, or stored in software in the memory of the aforementioned device, so that the processor can call and execute the operations corresponding to each module. The aforementioned device can be, but is not limited to, various types of data processing computer devices already existing in the art.

[0103] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0104] It is understood that, in addition to the memory and processor mentioned above, the computer equipment described above also includes other hardware and software components not listed in this specification. The specific components can be determined according to the model of the image processing computer in different application scenarios, and will not be listed and described in detail in this specification.

[0105] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0106] The above embodiments are merely illustrative of several implementation methods of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of protection of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and all such modifications and improvements fall within the scope of protection of this application.

Claims

1. A method for recognizing a chemical experiment device based on multi-layer feature fusion, characterized in that, Including the following steps: Step 1: Acquire RGB images of the chemical experiment scene and preprocess them to obtain input features; Step 2: Use a backbone network to extract multi-scale features from the input features to obtain multi-scale features; Step 3: Perform dynamic channel calibration on the multi-scale features using a channel attention module to obtain the channel-calibrated multi-scale features; Step 4: In the neck region, multi-branch parallel processing is used to perform deep enhancement and fusion of the multi-scale features after channel calibration to obtain multi-scale fused features. Specifically, this includes: the upper branch divides the multi-scale features after channel calibration into two paths: convolutional mapping and channel attention module calibration. After refining local features through the C2f module, the fusion is completed through the ADD module; the middle branch combines skip connections to complementaryly fuse shallow detail features and deep semantic features after processing through convolution, channel attention module, and C2f module; the lower branch upsamples deep small-size features through a double deconvolution module to achieve size alignment of multi-scale features; the channel attention module is used to dynamically calibrate channel weights, suppress redundant background information, and highlight key feature channels of the chemical experimental apparatus. Step 5: Input the multi-scale fusion features into the decoder head to obtain the chemical experimental device recognition result; The neck region comprises: seven convolutional modules, six channel attention modules, two deconvolutional modules, four C2f modules, and four ADD modules; within the neck region: After the small-sized features with channel calibration are processed by the first convolution module, the first convolutional features are obtained. The first convolutional feature is processed by the first channel attention module for dynamic channel calibration and then by the second convolutional module to obtain the second convolutional feature. The first convolutional feature is upsampled through the first deconvolution module to obtain the first upsampled feature; The first upsampled feature is input into the second channel attention module. The obtained channel attention feature is multiplied by the channel-calibrated mesoscale feature mapped by the third convolution module. The result is then input into the first ADD module along with the first upsampled feature to obtain the first fusion feature. The first fused feature is processed by the first C2f module and then dynamically calibrated by the third channel attention module. Finally, it is processed by the fourth convolution module to obtain the third convolution feature. The first upsampled feature is upsampled using a second deconvolution module to obtain the second upsampled feature; The second upsampled feature is input into the fourth channel attention module. The obtained channel attention feature is multiplied by the channel-calibrated large-size feature after processing by the fifth convolution module. The result and the second upsampled feature are then input into the second ADD module to obtain the second fusion feature. The second fusion feature is input into the second C2f module to obtain the mesoscale fusion feature; The mesoscale fusion features are input into the sixth convolutional module to obtain the fourth convolutional features; The fourth convolutional feature is input into the fifth channel attention module. The resulting attention feature is multiplied by the third convolutional feature. The result and the fourth convolutional feature are input into the third ADD module. The result is then input into the third C2f module to obtain the large-scale fusion feature. The fourth convolutional feature is input into the seventh convolutional module to obtain the fifth convolutional feature; The fifth convolutional feature is input into the sixth channel attention module. The obtained channel attention feature is multiplied by the second convolutional feature. The result and the fifth convolutional feature are input into the fourth ADD module. The result is input into the fourth C2f module to obtain the small-scale fusion feature.

2. The chemical experimental apparatus identification method based on multi-layer feature fusion according to claim 1, characterized in that, The backbone network mentioned in step 2 is the backbone network of the YOLOv8 network model.

3. The chemical experimental apparatus identification method based on multi-layer feature fusion according to claim 1, characterized in that, The channel attention module includes: a global max pooling layer, a global average pooling layer, and a sigmoid function; Step 3, which involves dynamic channel calibration using the channel attention module, includes the following steps: The features input to the channel attention module are processed by the global max pooling layer and the global average pooling layer respectively, and then added and fused. The fused result is activated by the Sigmoid function to obtain the channel attention features. The channel attention features are processed by channel-by-channel multiplication with the features input to the channel attention module to obtain the channel-calibrated features.

4. The chemical experimental apparatus identification method based on multi-layer feature fusion according to claim 1, characterized in that, The convolution module includes convolutional layers, batch normalized layers, and the SiLU activation function.

5. The chemical experimental apparatus identification method based on multi-layer feature fusion according to claim 1, characterized in that, The decoder head is used to further calibrate the feature channel weights using convolutional layers and a lightweight feature refinement module, enhancing the distinction between instrument semantic information and background. By detecting branches or segmenting branches, it outputs the category label and precise bounding box coordinates of the chemical experimental apparatus.

6. A chemical experimental apparatus identification device based on multi-layer feature fusion, characterized in that, include: The input feature determination unit is used to acquire RGB images of a chemical experiment scene and perform preprocessing to obtain input features; A multi-scale feature extraction unit is used to extract multi-scale features from the input features using a backbone network to obtain multi-scale features. The channel calibration unit is used to perform dynamic channel calibration on the multi-scale features using the channel attention module to obtain the channel-calibrated multi-scale features. A multi-scale feature fusion unit is used in the neck region to perform deep enhancement and fusion of channel-calibrated multi-scale features through multi-branch parallel processing to obtain multi-scale fused features. Specifically, the upper branch divides the channel-calibrated multi-scale features into two paths: convolutional mapping and channel attention module calibration. After refining local features by the C2f module, the fusion is completed by the ADD module. The middle branch combines skip connections to complementaryly fuse shallow detail features and deep semantic features after processing by convolution, the channel attention module, and the C2f module. The lower branch upsamples deep small-size features through a double deconvolution module to achieve size alignment of multi-scale features. The channel attention module is used to dynamically calibrate channel weights, suppress redundant background information, and highlight key feature channels of the chemical experimental apparatus. A chemical experimental apparatus identification unit is used to input the multi-scale fusion features into the decoder head to obtain the chemical experimental apparatus identification result; The neck region comprises seven convolutional modules, six channel attention modules, two deconvolutional modules, four C2f modules, and four ADD modules. Within the neck region: small-sized features after channel calibration are processed by the first convolutional module to obtain a first convolutional feature; the first convolutional feature is processed by the first channel attention module for dynamic channel calibration, then by the second convolutional module to obtain a second convolutional feature; the first convolutional feature is upsampled by the first deconvolutional module to obtain a first upsampled feature; the first upsampled feature is input into the second channel attention module, and the resulting channel attention feature is multiplied by the channel-calibrated mesoscale feature mapped by the third convolutional module. The result is then input into the first ADD module along with the first upsampled feature to obtain a first fused feature; the first fused feature is processed by the first C2f module, then dynamically calibrated by the third channel attention module, and then processed by the fourth convolutional module to obtain a third convolutional feature; the first upsampled feature is upsampled by the second deconvolutional module to obtain a second upsampled feature; the second upsampled feature is then... The channel attention feature is input into the fourth channel attention module, where it is multiplied by the channel-calibrated large-size feature processed by the fifth convolution module. The result and the second upsampled feature are then input into the second ADD module to obtain the second fusion feature. The second fusion feature is input into the second C2f module to obtain the mesoscale fusion feature. The mesoscale fusion feature is input into the sixth convolution module to obtain the fourth convolution feature. The fourth convolution feature is input into the fifth channel attention module, where it is multiplied by the third convolution feature. The result and the fourth convolution feature are then input into the third ADD module, and the result is input into the third C2f module to obtain the large-scale fusion feature. The fourth convolution feature is input into the seventh convolution module to obtain the fifth convolution feature. The fifth convolution feature is input into the sixth channel attention module, where it is multiplied by the second convolution feature. The result and the fifth convolution feature are then input into the fourth ADD module, and the result is input into the fourth C2f module to obtain the small-scale fusion feature.

7. The chemical experimental apparatus identification device based on multi-layer feature fusion according to claim 6, characterized in that, The backbone network in the multi-scale feature extraction unit is the backbone network of the YOLOv8 network model.

8. The chemical experimental apparatus identification device based on multi-layer feature fusion according to claim 6, characterized in that, The channel attention module includes a global max pooling layer, a global average pooling layer, and a sigmoid function. The specific process of dynamic channel calibration using the channel attention module in the channel calibration unit includes: processing the features input to the channel attention module through the global max pooling layer and the global average pooling layer respectively, then adding and fusing them; activating the fused result using the sigmoid function to obtain the channel attention features; and processing the channel attention features and the features input to the channel attention module through a channel-by-channel multiplication operation to obtain the channel-calibrated features.

9. A computer device, comprising a memory and a processor, characterized in that, The memory stores a computer program, and when the processor executes the computer program, it implements the steps of the chemical experimental apparatus identification method based on multi-layer feature fusion as described in any one of claims 1 to 5.

Images