Light-weight instrument small target detection model and method for complex industrial scene

Abstract

The invention discloses a light-weight instrument small target detection model and method for a complex industrial scene, belongs to the crossing field of deep learning and edge calculation, and aims to solve the problem that an existing method cannot meet the real-time detection requirements of edge equipment such as an inspection robot in the aspects of precision, efficiency and small target detection capability. The model comprises a lightweight backbone network used for extracting multi-scale features from an input image; the cross-scale feature fusion network is used for bidirectionally fusing the multi-scale features, retaining shallow space details and deep semantic information and outputting fused features; the deformable large-kernel target sensing module is deployed at a specified position of the cross-scale feature fusion network so as to better capture feature information related to a target area and improve the feature expression capability of a small target; and the multi-task decoupling prediction network performs classification, positioning regression and confidence prediction on the input image in parallel, and a positioning regression branch adopts an EIOU loss function of decoupling width and height optimization.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of deep learning and edge computing, specifically involving a lightweight instrument small target detection method for complex industrial scenarios, which is particularly suitable for embedded deployment of substation inspection robots. Background Technology

[0002] In industrial substations, petrochemical plants, and rail transportation, numerous pointer-type or digital instruments are used to monitor key parameters such as pressure, temperature, current, and voltage in real time. Traditional manual inspection methods rely on human visual readings, which suffer from inherent drawbacks such as low efficiency, large subjective errors, and poor reliability at night or in inclement weather. With the widespread application of intelligent inspection robots, automated instrument detection and reading based on computer vision has become a research hotspot.

[0003] Inspection robots are typical edge computing devices, and their embedded hardware (such as Jetson Nano and RK3588NPU) is characterized by limited computing power (typically ≤5 TOPS), limited memory resources (4–8GB RAM), and power sensitivity (≤15W). Therefore, the instrument detection algorithms deployed on inspection robots must meet stringent requirements of lightweight design, high accuracy, and real-time performance. However, existing technologies still face many challenges in edge computing scenarios.

[0004] Currently, computer vision-based instrument target detection methods are mainly divided into two categories:

[0005] 1. Two-stage detection method

[0006] Two-stage methods, such as Faster R-CNN and Mask R-CNN, typically consist of two steps: candidate region generation and object detection. First, a region proposal network or similar mechanism generates candidate boxes that may contain the target. Then, feature extraction and classification / regression are performed within these candidate regions. Some studies using this type of model have achieved high detection accuracy by optimizing feature extraction, region proposal, and detection network architecture. However, these methods are computationally complex and slow inference, typically requiring significant computational resources, making them unsuitable for deployment on edge devices or in real-time industrial scenarios.

[0007] 2. Single-stage detection method

[0008] Single-stage methods, such as the YOLO series, directly regress bounding boxes and categories on feature maps, thus significantly improving detection speed. They also have a relatively simple structure, making them easy to port to computationally limited edge devices. However, in complex background and small object detection tasks, single-stage models often suffer from accuracy degradation.

[0009] Given the limitations of existing technologies, industrial instrumentation in edge computing scenarios faces the following core challenges:

[0010] The contradiction between accuracy and efficiency: high-precision models require computing power that exceeds the capabilities of edge devices, while lightweight models lack sufficient accuracy.

[0011] Small target detection fails: Traditional multi-scale fusion methods lose information about small targets in deep features, leading to a surge in false negative rates.

[0012] Reliance on high-quality images: Previous studies have struggled to adapt to low-quality images captured by inspection robots in complex scenarios.

[0013] Positioning error: Most studies focus on the success rate of detection, while ignoring the precise positioning of the target detection box, which leads to insufficient accuracy in subsequent perspective distortion correction and affects the accuracy of readings.

[0014] It is evident that existing methods fail to meet the real-time detection requirements of edge devices such as inspection robots in terms of accuracy, efficiency, and small target detection capabilities. Therefore, there is an urgent need for a lightweight, high-precision instrument detection algorithm that can achieve efficient and stable operation on resource-constrained hardware, while simultaneously addressing core issues such as missed detection of small targets and inaccurate positioning. Summary of the Invention

[0015] To address the problem that existing methods cannot meet the real-time detection requirements of edge devices such as inspection robots in terms of accuracy, efficiency, and small target detection capabilities, this invention provides a lightweight instrument small target detection model and method for complex industrial scenarios.

[0016] In one aspect, this invention provides a lightweight instrument small target detection model for complex industrial scenarios, including a lightweight backbone network, a cross-scale feature fusion network, a deformable large-kernel target perception module, and a multi-task decoupled prediction network.

[0017] Lightweight backbone network: used to extract multi-scale features from input images with limited computing resources;

[0018] Cross-scale feature fusion network: connected to the lightweight backbone network, used to bidirectionally fuse the multi-scale features, retaining shallow spatial details and deep semantic information, and outputting fused features;

[0019] Deformable large-kernel target perception module: Deployed at designated locations in the cross-scale feature fusion network to better capture feature information related to the target region, including:

[0020] Global Context Extraction Unit: A large receptive field is constructed using depthwise separable convolution and dilated convolution to capture multi-scale contextual information;

[0021] Adaptive Spatial Sampling Unit: Learns spatial offset through deformable convolution to dynamically adjust feature sampling position;

[0022] Feature Enhancement Output Unit: Integrates global context and adaptive sampling results to highlight the target region and suppress background interference, thereby improving the feature representation capability of small targets;

[0023] The multi-task decoupled prediction network performs classification, localization regression, and confidence prediction in parallel on the input image. The localization regression branch uses the decoupled width and height optimized EIOU loss function.

[0024] Preferably, the lightweight backbone network includes:

[0025] Standard convolutional layers are used to perform initial downsampling of the input image;

[0026] Multiple inverted residual bottleneck modules, or Bneck modules, are stacked sequentially to form at least four stages to achieve multi-level downsampling;

[0027] The Spatial Pyramid Pooling Module (SPPF) is used to integrate global context information;

[0028] The lightweight backbone network outputs a shallow feature map P3 with a downsampling rate of 1 / 8, a mid-level feature map P4 with a downsampling rate of 1 / 16, and a deep feature map P5 with a downsampling rate of 1 / 32.

[0029] Preferably, the standard convolutional layer of the lightweight backbone network is a 3×3 convolutional layer, and the 3×3 convolutional layer has 16 output channels and a stride of 2.

[0030] The four stages of the Bneck modules stacked sequentially are:

[0031] Phase 1: 2 Bneck modules. Each Bneck module in this phase contains a 3×3 depthwise separable convolution with 24 output channels.

[0032] Second stage: 3 Bneck modules. In this stage, the Bneck modules contain 5×5 depthwise separable convolutions with 40 output channels.

[0033] The third stage consists of two Bneck modules, each containing a 5×5 depthwise separable convolution with 48 output channels.

[0034] Fourth stage: 3 Bneck modules. The Bneck modules in this stage contain 5×5 depthwise separable convolutions with 96 output channels.

[0035] Then, after the feature compression layer, the number of channels is adjusted to the required dimension and then sent to the Spatial Pyramid Pooling Module (SPPF) for multi-scale pooling and global context integration.

[0036] Feature compression layer: 1×1 convolution, the 1×1 convolution output channel 576;

[0037] The output channel of the Spatial Pyramid Pooling Module (SPPF) is 1024.

[0038] Preferably, the cross-scale feature fusion network includes two bidirectional fusion paths:

[0039] First fusion path: The spatial resolution of the deep low-resolution feature map is gradually increased by upsampling units to align with the shallow high-resolution feature map, and then the feature maps are stitched together to form the first fusion path from top to bottom.

[0040] The second fusion path: The spatial resolution of the shallow high-resolution feature map is gradually reduced by the downsampling unit to align with the deep low-resolution feature map, and then the feature maps are stitched together to form a bottom-up second fusion path.

[0041] The cross-scale feature fusion network outputs fused features corresponding to the three detection scales P3, P4, and P5, respectively.

[0042] Preferably, the first fusion path of the cross-scale feature fusion network includes two upsampling units, using a cross-stage local double convolutional layer C2f (abbreviation of CSPlayer_2Conv) module for fusion; the first upsampling unit upsamples the deep feature map P5 by a factor of 2 to obtain a first upsampled feature with the same spatial size as the middle feature map P4; the first upsampled feature is concatenated with the middle feature map P4 along the channel dimension and fed into the C2f module for fusion to generate a first fused feature F4; the second upsampling unit further upsamples F4 by a factor of 2 to obtain a second upsampled feature with the same spatial size as the shallow feature map P3; the second upsampled feature is concatenated with the shallow feature map P3 along the channel dimension and fed into the C2f module to generate a second fused feature F3;

[0043] The first fusion path is set with a first convolutional compression unit, which is used to perform 1×1 convolution on F3 and F4 in the first fusion path respectively to reduce the number of channels;

[0044] The second fusion path of the cross-scale feature fusion network includes a downsampling unit, which uses the C2f module for fusion. A bottom-up reverse path is constructed using the downsampling unit, where the channel-compressed F3 is downsampled with a step size of 2. After downsampling, F3 is concatenated with the channel-compressed F4 and fed into the C2f module to generate the third fusion feature F5. The third fusion feature F5 is then downsampled with a step size of 2 and concatenated with the deep feature map P5, fed into the C2f module to generate the fourth fusion feature F6.

[0045] The second fusion path includes a second convolutional compression unit, which performs a 1×1 convolution on the third fusion feature F5 to unify the channel dimensions.

[0046] Preferably, the two fusion paths of the cross-scale feature fusion network are respectively provided with feature input ends and output ends;

[0047] At the feature input end, three 1×1 convolutions are used to receive shallow feature maps P3 (1 / 8 downsampling rate), mid-layer feature maps P4 (1 / 16 downsampling rate), and deep feature maps P5 (1 / 32 downsampling rate) from the lightweight backbone network, respectively.

[0048] At the output end, three final fused feature maps are output: F3_out (1 / 8 resolution), F4_out (1 / 16 resolution), and F5_out (1 / 32 resolution), which are used by the decoupled prediction head.

[0049] Preferably, the deformable large-kernel target perception module is set at a designated node of the cross-scale feature fusion network, wherein the designated node is:

[0050] In the first fusion path, F4 is the first feature node after upsampling and fusion with P3; and

[0051] In the second fusion path, the second feature node is obtained by splicing and fusing F3 after downsampling with F4 after channel compression;

[0052] The deformable large kernel target perception module is implemented using a deformable large kernel attention mechanism, which is applied to the fused feature maps of the first feature node and the second feature node, respectively.

[0053] Preferably, the deformable large-kernel target sensing module includes:

[0054] A deformable convolution submodule is used to dynamically adjust the convolution sampling position based on the offset matrix;

[0055] A multi-scale large-kernel convolutional submodule is used to extract contextual features from different receptive fields in parallel.

[0056] The attention-weighted submodule is used to apply attention weights to the input feature map through element-wise multiplication to enhance the feature representation of small target regions.

[0057] In another aspect, the present invention provides a lightweight instrument small target detection method for complex industrial scenarios, the method comprising the following steps:

[0058] Step 1: Construct an image dataset containing small-sized instrument targets, and re-annotate and augment the images in the dataset;

[0059] Step 2: Input the enhanced dataset into the object detection model for object recognition training, where:

[0060] A lightweight backbone network receives the original image, downsamples it through a multi-level inverse residual bottleneck module Bneck, and outputs features P3, P4, and P5 at three scales: 1 / 8, 1 / 16, and 1 / 32, respectively.

[0061] The cross-scale feature fusion network first passes P3, P4, and P5 through a 1×1 convolution to unify the channels, and then performs the following sequentially:

[0062] Top-down approach: Upsampling - concatenation - fusion - convolutional compression to generate fused features F3 and F4;

[0063] Bottom-up approach: downsampling - concatenation - fusion - convolutional compression to generate fused features F5 and F6;

[0064] The final output consists of three fused features: F3_out, F4_out, and F5_out.

[0065] A deformable large kernel target perception module is inserted after F3_out and F4_out to enhance the feature representation of small target regions.

[0066] The multi-task decoupled prediction network feeds three fused features, F3_out, F4_out, and F5_out, into the network in parallel. The classification branch outputs the class probability, and the regression branch outputs the center point, width, and height. After optimization by the EIOU loss function, the final detection box and class are obtained, and the class, location, and confidence recognition results are output.

[0067] Step 3: Deploy the trained target detection model on an edge device with limited computing resources to perform instrument target detection tasks.

[0068] Preferably, the edge device is an inspection robot.

[0069] The beneficial effects of this invention are as follows: This invention addresses the perspective images from the inspection robot in complex industrial environments such as substations, proposing a lightweight and high-precision method for detecting small targets on instruments. Through four innovations—a lightweight backbone network, cross-scale feature fusion, a deformable large-kernel attention mechanism, and accurate regression—it achieves a perfect balance between high precision (mAP50 reaching 91.36%) and strong robustness (small target recall rate of 90.62%) on edge computing devices for the first time. Many technical indicators surpass existing solutions, providing a reliable edge computing solution for intelligent inspection of industrial instruments and possessing significant industrial application value. Specifically, this is reflected in the following aspects:

[0070] 1. In terms of balancing accuracy and efficiency, a lightweight backbone network is adopted. Through an optimized combination of depthwise separable convolution, inverse residual structure, and h-swish activation function, the computational complexity is significantly reduced while maintaining detection accuracy. Experimental data shows that the computational cost of this model is reduced to 17.7 GFLOPs, while the model size is compressed to 14.4 MB. The reduced computational cost and smaller model size make this model more suitable for deployment on edge devices with limited computing resources.

[0071] 2. To address the issue of failure in small target detection, an innovative cross-scale feature fusion (CNN-based) algorithm is incorporated. This network effectively enhances the coupling between shallow detail features and deep semantic features by constructing a bidirectional, multi-scale feature interaction path, improving the model's detection capabilities for small targets, occluded targets, and instrument regions in complex industrial backgrounds, thereby enhancing the model's generalization ability and robustness. Test results show that the recall rate for small instrument targets is improved to 90.62%.

[0072] 3. In terms of improving positioning accuracy, the EIOU loss function is adopted to decouple the calculation of width and height loss and introduce the minimum closed rectangle as a normalization factor, which improves the accuracy and convergence speed of the model for bounding box positioning.

[0073] 4. To address the limitations of a fixed receptive field, this patent innovatively introduces a deformable large kernel attention mechanism (D-LKA). This mechanism dynamically adjusts the feature sampling position through deformable convolution, and, in conjunction with multi-scale large kernel convolution to extract contextual information, improves the accuracy by 2.33%, effectively overcoming the feature extraction bias of traditional convolution under instrument perspective deformation.

[0074] Based on the above advantages, the technology of this invention has great engineering application value. The 14.4MB model size of this invention can be perfectly adapted to edge device deployment, meeting the requirements of long-term operation of inspection robots. The overall model has the advantages of superior accuracy and high computing efficiency, and can be adapted to resource-constrained inspection robot platforms to achieve efficient detection and positioning of key components in industrial dashboards, showing good prospects for engineering application and promotion. Attached Figure Description

[0075] Figure 1 This is a schematic diagram of the structure of a lightweight instrument small target detection model for complex industrial scenarios as described in this invention;

[0076] Figure 2 This is a flowchart of a lightweight instrument small target detection method for complex industrial scenarios as described in this invention;

[0077] Figure 3 This is the schematic diagram of the inverted residual bottleneck module Bneck;

[0078] Figure 4 This is a schematic diagram of the CCFF algorithm;

[0079] Figure 5 This is a schematic diagram illustrating the principle of introducing a deformable large kernel attention mechanism into a cross-scale feature fusion network.

[0080] Figure 6 This is a schematic diagram of the deformable large kernel attention mechanism;

[0081] Figure 7 It is a comparison of the dataset before and after the annotation was modified, where Figure 7 (a) is the image before data A was labeled. Figure 7 (b) is the image after data A has been labeled; Figure 7 (c) is the image before data B was labeled. Figure 7 (d) is the image after data B is labeled;

[0082] Figure 8 It is an added dataset, in which Figure 8 (a) is a photograph of the original pointer-type instrument. Figure 8 (b) A representative image selected from 40 digital meters from Roboflow;

[0083] Figure 9 This is a diagram illustrating data augmentation;

[0084] Figure 10 This is a heatmap comparison before and after the introduction of the deformable large kernel attention mechanism, in which... Figure 10 (a) is the original image. Figure 10 (b) Heatmap before the introduction of the deformable large kernel attention mechanism. Figure 10 (c) Heatmap after introducing deformable macronucleus attention mechanism;

[0085] Figure 11 This is a schematic diagram of some results obtained by the instrument using the model of this invention;

[0086] Figure 12 This is a comparison chart of the results of the detection instrument using the model of this invention and the existing YOLOv8s model. The left column shows the results of the detection instrument using the existing YOLOv8s model, and the right column shows the results of the detection instrument using the model of this invention. Detailed Implementation

[0087] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0088] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0089] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but this is not intended to limit the scope of the invention.

[0090] Specific Implementation Method 1: The following is combined with... Figures 1 to 12 This embodiment describes a lightweight instrument small target detection model for complex industrial scenarios, which includes a lightweight backbone network, a cross-scale feature fusion network, a deformable large-kernel target perception module, and a multi-task decoupled prediction network.

[0091] Lightweight backbone network: used to extract multi-scale features from input images with limited computing resources;

[0092] Cross-scale feature fusion network: connected to the lightweight backbone network, used to bidirectionally fuse the multi-scale features, retaining shallow spatial details and deep semantic information, and outputting fused features;

[0093] Deformable large-kernel target perception module: Deployed at designated locations in the cross-scale feature fusion network to better capture feature information related to the target region, including:

[0094] Global Context Extraction Unit: A large receptive field is constructed using depthwise separable convolution and dilated convolution to capture multi-scale contextual information;

[0095] Adaptive Spatial Sampling Unit: Learns spatial offset through deformable convolution to dynamically adjust feature sampling position;

[0096] Feature Enhancement Output Unit: Integrates global context and adaptive sampling results to highlight the target region and suppress background interference, thereby improving the feature representation capability of small targets;

[0097] The multi-task decoupled prediction network performs classification, localization regression, and confidence prediction in parallel on the input image. The localization regression branch uses the decoupled width and height optimized EIOU loss function.

[0098] See Figure 1The lightweight instrument small target detection model for complex industrial scenarios comprises three main parts: a lightweight backbone network, a cross-scale feature fusion network, and a multi-task decoupled prediction network. The cross-scale feature fusion network embeds a deformable large-kernel target perception module.

[0099] Lightweight backbone networks include:

[0100] Standard convolutional layers are used to perform initial downsampling of the input image;

[0101] Multiple inverted residual bottleneck modules, or Bneck modules, are stacked sequentially to form at least four stages to achieve multi-level downsampling;

[0102] The Spatial Pyramid Pooling Module (SPPF) is used to integrate global context information;

[0103] The lightweight backbone network outputs a shallow feature map P3 with a downsampling rate of 1 / 8, a mid-level feature map P4 with a downsampling rate of 1 / 16, and a deep feature map P5 with a downsampling rate of 1 / 32.

[0104] Specifically, the structure of the lightweight backbone network is as follows:

[0105] First, the standard convolutional layer uses a 3×3 convolutional layer, which has 16 output channels and a stride of 2.

[0106] Then the Bneck modules are stacked in four stages in sequence:

[0107] Phase 1: 2 Bneck modules. Each Bneck module in this phase contains a 3×3 depthwise separable convolution with 24 output channels.

[0108] Second stage: 3 Bneck modules. In this stage, the Bneck modules contain 5×5 depthwise separable convolutions with 40 output channels.

[0109] The third stage consists of two Bneck modules, each containing a 5×5 depthwise separable convolution with 48 output channels.

[0110] Fourth stage: 3 Bneck modules. The Bneck modules in this stage contain 5×5 depthwise separable convolutions with 96 output channels.

[0111] Finally, the number of channels is adjusted to the required dimension by the feature compression layer and then sent to the Spatial Pyramid Pooling Module (SPPF) for multi-scale pooling and global context integration.

[0112] Feature compression layer: 1×1 convolution, the 1×1 convolution output channel 576;

[0113] The output channel of the Spatial Pyramid Pooling Module (SPPF) is 1024.

[0114] The backbone network ultimately outputs feature maps at multiple scales, corresponding to downsampling rates of 1 / 8, 1 / 16, and 1 / 32 of the input image, providing rich and multi-layered semantic feature support for subsequent feature fusion and detection tasks.

[0115] See Figure 3 The Bneck module structure is the core lightweight building block in the backbone network, and its full name is Inverted Residual Bottleneck. Its design goal is to significantly reduce computational load while maintaining accuracy, making it particularly suitable for deployment on edge devices.

[0116] The Bneck module first performs dimensionality upscaling for feature extraction, then dimensionality reduction. Dimensionality upscaling expands the input channel count by 4-6 times using 1×1 convolutions, performing feature transformation in a higher-dimensional space to enhance expressive power. Next, it extracts features using depthwise separable convolutions with 3×3 or 5×5 kernels, performing independent convolutions channel by channel. SE attention-based global average pooling is used to compress spatial information, dynamically improving the response of key instrument features. Finally, 1×1 convolutions compress the channels back to the target dimension, employing linear activation to avoid information loss.

[0117] The Bneck module combines inverted residual structure, depthwise separable convolution, and SE attention (Squeeze-and-Excitation) into a single, highly effective approach. The SE attention module dynamically enhances key channels by compressing and stimulating the module, focusing attention on the instrument region. This reduces computational load by 37.7% while improving the detection performance of small targets on instruments in complex scenarios. Its structured deployment (specific configuration of channel number / convolution kernel / stage number) in the patent is a key technological support for achieving edge detection of instruments.

[0118] This backbone network combines hardware-aware network architecture search (NAS) and the NetAdapt algorithm to achieve more refined structure optimization. At the same time, by improving the inverse residual bottleneck structure, introducing a lightweight SE module, an efficient hard swish activation function, and redesigning the computationally expensive layers in the network, the computational cost and latency of the model are significantly reduced, effectively achieving network lightweighting while improving accuracy.

[0119] The cross-scale feature fusion network includes two bidirectional fusion paths, with feature input and output terminals set before and after each path, respectively.

[0120] At the feature input end, three 1×1 convolutions are used to receive shallow feature maps P3 (1 / 8 downsampling rate), mid-layer feature maps P4 (1 / 16 downsampling rate), and deep feature maps P5 (1 / 32 downsampling rate) from the lightweight backbone network, respectively.

[0121] First fusion path: The spatial resolution of the deep low-resolution feature map is gradually increased by upsampling units to align with the shallow high-resolution feature map, and then the feature maps are stitched together to form the first fusion path from top to bottom.

[0122] The second fusion path: The spatial resolution of the shallow high-resolution feature map is gradually reduced by the downsampling unit to align with the deep low-resolution feature map, and then the feature maps are stitched together to form a bottom-up second fusion path.

[0123] At the output end, three final fused feature maps are output: F3_out (1 / 8 resolution), F4_out (1 / 16 resolution), and F5_out (1 / 32 resolution), which are used by the decoupled prediction head.

[0124] The cross-scale feature fusion network outputs fused features corresponding to the three detection scales P3, P4, and P5, respectively.

[0125] The first fusion path of the cross-scale feature fusion network includes two upsampling units, which are fused using a C2f module. The first upsampling unit upsamples the deep feature map P5 by a factor of 2 to obtain a first upsampled feature with the same spatial size as the middle feature map P4. The first upsampled feature is then concatenated with the middle feature map P4 along the channel dimension and fed into the C2f module for fusion, generating a first fused feature F4. The second upsampling unit further upsamples F4 by a factor of 2 to obtain a second upsampled feature with the same spatial size as the shallow feature map P3. The second upsampled feature is then concatenated with the shallow feature map P3 along the channel dimension and fed into the C2f module to generate a second fused feature F3.

[0126] The first fusion path is set with a first convolutional compression unit, which is used to perform 1×1 convolution on F3 and F4 in the first fusion path respectively to reduce the number of channels;

[0127] The second fusion path of the cross-scale feature fusion network includes a downsampling unit, which uses the C2f module for fusion. A bottom-up reverse path is constructed using the downsampling unit, where the channel-compressed F3 is downsampled with a step size of 2. After downsampling, F3 is concatenated with the channel-compressed F4 and fed into the C2f module to generate the third fusion feature F5. The third fusion feature F5 is then downsampled with a step size of 2 and concatenated with the deep feature map P5, fed into the C2f module to generate the fourth fusion feature F6.

[0128] The second fusion path sets up a second convolutional compression unit, which is used to perform a 1×1 convolution on the third fusion feature F5 to unify the channel dimension;

[0129] As can be seen, the idea behind the cross-scale feature fusion network fusion CCFF (CNN-based Cross-scale Feature Fusion) algorithm can be found in [link to documentation]. Figure 4 As shown, the CCFF algorithm performs bidirectional fusion of feature maps at different scales, preserving detailed information of low-resolution features; it uses 1x1 convolutions to reduce the number of channels in the fused features, thus reducing model complexity. Multiple scale feature maps are extracted from the backbone network, namely shallow feature map P3, mid-level feature map P4, and deep feature map P5; P3, P4, and P5 feature maps are compressed using 1×1 convolutions to unify their output channel count; the deep feature map P5 is upsampled and concatenated with the mid-level feature map P4, and the concatenation result is fed into a C2f convolution for fusion, generating fused feature F4; F4 is further upsampled, concatenated with the shallow feature map P3, and fed into the same double convolution structure for further fusion, generating fused feature F3; Cross-scale convolutional fusion modules are added to the F4 and F3 fusion paths respectively. Local contextual information is extracted through multi-layer 1×1 and 3×3 convolutions to enhance the interactive expression between multi-scale features. For the fused features such as F3 and F4, a bottom-up reverse path is constructed through downsampling operation to supplement the low-level edge information in the high-level features and realize bidirectional information flow. The final multi-scale fused feature maps are output, corresponding to three detection scales of 1 / 8, 1 / 16 and 1 / 32 respectively, as input features of the multi-task decoupling prediction network.

[0130] This invention also embeds a deformable large kernel target perception module, namely the Deformable Large Kernel Attention (D-LKA) mechanism, into the cross-scale feature fusion network. See the schematic diagram of the principle of embedding the Deformable Large Kernel Attention mechanism into the cross-scale feature fusion network. Figure 5 As shown, large convolutional kernels are used to capture long-distance contextual features. Deformable convolutions are used to dynamically adjust the feature sampling positions, improving the model's sensitivity to small targets in complex backgrounds. The deformable large-kernel target perception module is set at a designated node in the cross-scale feature fusion network, and the designated node is:

[0131] In the first fusion path, the first feature node after F4 is upsampled and concatenated with P3 (after F3_out); and

[0132] In the second fusion path, the second feature node (after F4_out) is obtained by concatenating and fusing F3 after downsampling with F4 after channel compression.

[0133] The deformable large kernel target perception module is implemented using a deformable large kernel attention mechanism, which is applied to the fused feature maps of the first feature node and the second feature node, respectively.

[0134] See Figure 6As shown, the deformable large-kernel target perception module includes:

[0135] A deformable convolution submodule is used to dynamically adjust the convolution sampling position based on the offset matrix;

[0136] A multi-scale large-kernel convolutional submodule is used to extract contextual features from different receptive fields in parallel.

[0137] The attention-weighted submodule is used to apply attention weights to the input feature map through element-wise multiplication to enhance the feature representation of small target regions.

[0138] In the D-LKA module, the input feature map is first spatially modeled using multi-scale deformable convolution. This structure learns the offsets of key sampling positions and dynamically adjusts the receptive field to adapt to target regions with different deformations and sizes in the image, thereby achieving adaptive enhancement and improved localization accuracy of key semantic regions. Subsequently, a multi-scale parallel large-kernel convolution structure is used to extract deep contextual information from the feature map, achieving complementary enhancement of local and global semantics through combinations of different receptive field sizes. This module integrates multi-scale feature responses, enhancing the model's ability to recognize difficult-to-detect targets such as blurred edges and small targets in complex backgrounds, effectively improving the network's semantic focusing performance and environmental robustness.

[0139] The multi-task decoupled prediction network performs classification, localization regression, and confidence prediction in parallel on the input image. The localization regression branch uses the decoupled width and height optimized EIOU loss function.

[0140] In the training process of the object detection model, the EIOU loss function is used to calculate the regression error between the predicted and ground truth boxes. The EIOU loss function measures the difference between the predicted and ground truth boxes from three dimensions: center distance error, width error, and height error. By explicitly separating the regression of these three geometric quantities, the model is more likely to converge. In the error calculation process, the EIOU loss function no longer uses the normalized diagonal distance as a penalty term, but instead directly regresses the width and height differences separately, thus avoiding the error weight shift problem dominated by the long side in CIOU. The EIOU loss is used as the main object box regression loss function in the regression branch, and together with the classification loss and confidence loss, it forms the final multi-task total loss function to optimize the model training performance. This loss function effectively improves the regression speed and accuracy, especially in scenarios with small objects and low bounding box overlap, it can more stably guide the predicted boxes to move closer to the ground truth boxes, further enhancing the model's robustness in detecting instrument targets in complex scenarios.

[0141] Finally, the model trained by this invention is deployed to an edge computing device (such as an inspection robot) for instrument target detection.

[0142] See Figure 2 As shown, the instrument target detection method based on the lightweight instrument target detection model for complex industrial scenarios includes the following steps:

[0143] Step 1: Construct an image dataset containing small-sized instrument targets, and re-annotate and augment the images in the dataset;

[0144] Step 2: Input the enhanced dataset into the object detection model for object recognition training, where:

[0145] A lightweight backbone network receives the original image, downsamples it through a multi-level inverse residual bottleneck module Bneck, and outputs features P3, P4, and P5 at three scales: 1 / 8, 1 / 16, and 1 / 32, respectively.

[0146] The cross-scale feature fusion network first passes P3, P4, and P5 through a 1×1 convolution to unify the channels, and then performs the following sequentially:

[0147] Top-down approach: Upsampling - concatenation - fusion - convolutional compression to generate fused features F3 and F4;

[0148] Bottom-up approach: downsampling - concatenation - fusion - convolutional compression to generate fused features F5 and F6;

[0149] The final output consists of three fused features: F3_out, F4_out, and F5_out.

[0150] A deformable large kernel target perception module is inserted after F3_out and F4_out to enhance the feature representation of small target regions.

[0151] The multi-task decoupled prediction network feeds three fused features, F3_out, F4_out, and F5_out, into the network in parallel. The classification branch outputs the class probability, and the regression branch outputs the center point, width, and height. After optimization by the EIOU loss function, the final detection box and class are obtained, and the class, location, and confidence recognition results are output.

[0152] Step 3: Deploy the trained target detection model on an edge device with limited computing resources to perform instrument target detection tasks.

[0153] The edge device is an inspection robot.

[0154] The method will be described below with a specific embodiment.

[0155] This example provides a lightweight instrument target detection method suitable for inspection robots. This method is based on a lightweight instrument target detection model for complex industrial scenarios, and includes the following steps:

[0156] Step 1: Construct an image dataset containing small-sized instrument targets, and re-annotate and augment the images in the dataset;

[0157] The image dataset is an instrument dataset specifically designed for the perspective of inspection robots in complex industrial settings such as substations.

[0158] This invention uses the Meter Challenge (MC1296) dataset, which contains 1296 images captured by an automated robot. Furthermore, to facilitate correction after dashboard detection, this invention re-annotated the original dataset, ensuring the detection bounding boxes are as tangent as possible to the dashboard outline. This also resolves the issues of mislabeling and omissions in the original annotations. The annotations before and after the modification are shown below. Figure 7 As shown. To further improve the localization capability of dial detection, 262 images from a pointer-type instrument dataset and 40 images from a digital instrument dataset from Roboflow were added to the MC1296 dataset, totaling 1674 images, as shown. Figure 8 As shown. To improve the model's adaptability to real-world scenarios and detection accuracy, this paper performs random data augmentation on these 1674 datasets, such as... Figure 9 As shown, this includes: blur, brightness adjustment, flip, rotation, and salt and pepper attack, such as... Figure 9 As shown. These data augmentations mainly cover richer lighting conditions and more diverse shooting angles, and the number of images has been expanded from 1674 to 5673, effectively reducing the risk of overfitting and enhancing the model's generalization ability. After all the above datasets were labeled, they were divided into a training set (80%) and a validation set (20%).

[0159] Step 2: Input the enhanced dataset into the object detection model for object recognition training, where:

[0160] A lightweight backbone network receives the original image, downsamples it through a multi-level inverse residual bottleneck module Bneck, and outputs features P3, P4, and P5 at three scales: 1 / 8, 1 / 16, and 1 / 32, respectively.

[0161] Throughout all the above stages, the Bneck module introduces a lightweight SE attention mechanism to dynamically enhance key features through channel attention; at the same time, it uses the h-swish activation function to replace the traditional ReLU, improving the model's computational efficiency and fitting ability on mobile devices.

[0162] After the Bneck modules are stacked, the feature map undergoes a 1×1 convolution, expanding the number of channels to 576 for unified integration of feature dimensions. The feature map after the 1×1 convolution is then input into the SPPF module in YOLOv8, where multi-scale pooling is used to integrate the features in context, ultimately outputting a deep semantic feature map with 1024 channels.

[0163] This backbone network ultimately generates feature maps with three downsampling rates, namely:

[0164] P3: Downsampling factor 1 / 8;

[0165] P4: Downsampling factor 1 / 16;

[0166] P5: Downsampling factor 1 / 32.

[0167] The aforementioned multi-scale feature maps will be input into the Neck structure for further feature fusion and detection processing.

[0168] The cross-scale feature fusion network first passes P3, P4, and P5 through a 1×1 convolution to unify the channels, and then performs the following sequentially:

[0169] Top-down approach: Upsampling - concatenation - fusion - convolutional compression to generate fused features F3 and F4;

[0170] Bottom-up approach: downsampling - concatenation - fusion - convolutional compression to generate fused features F5 and F6;

[0171] This embodiment incorporates the CCFF algorithm concept, such as... Figure 9 As shown, this enhances the interaction capabilities between multi-scale features, effectively improving the detection performance for complex backgrounds, small-sized targets, and occluded targets, while reducing computational complexity.

[0172] Three feature maps at different scales are obtained from the backbone network, corresponding to feature maps with downsampling rates of 1 / 8, 1 / 16, and 1 / 32 of the input image, respectively. These three scales are selected as inputs to the cross-scale feature fusion network, denoted as P3, P4, and P5. The high-level feature map P5 (1 / 32 scale) is upsampled, doubling its spatial size to achieve spatial alignment with the feature map P4 (1 / 16 scale). Subsequently, the upsampled P5 and P4 are concatenated along the channel dimension to generate the first fused feature F4. Similarly, F4 is upsampled again and concatenated with P3 (1 / 8 scale) to obtain the final fused feature, the second fused feature F3.

[0173] On the stitched fused feature map, standard 1×1 and 3×3 convolutions are introduced to compress the channel dimensions and enhance the local receptive field. This module not only integrates semantic information from different levels but also preserves detailed edge features. Specifically, the 1×1 convolution is used for channel compression to reduce redundant information and computational overhead, while the 3×3 convolution further integrates spatial local contextual information.

[0174] The fused F3 is passed up to higher levels to construct a cascaded feature fusion path. In each level of feature fusion, an upsampling + concatenation + convolution enhancement strategy is employed to ensure that semantic information is progressively passed down to lower levels while preserving spatial details. The channel-compressed F3 is downsampled with a stride of 2. After downsampling, F3 is concatenated with the channel-compressed F4 and fed into the C2f module to generate the third fused feature F5. The third fused feature F5 is then downsampled with a stride of 2 and concatenated with the deep feature map P5 before being fed into the C2f module to generate the fourth fused feature F6.

[0175] The final output consists of three fused features: F3_out, F4_out, and F5_out; F3_out = F3, F4_out = F5, and F5_out = F6. These features are used in different branches of the subsequent multi-task decoupling prediction network to perform accurate identification of targets of different sizes.

[0176] The CCFF algorithm, based on cross-scale feature fusion networks, significantly enhances the model's target representation capabilities in complex industrial scenarios. It addresses the semantic weakness and spatial distortion issues inherent in traditional FPN+PAN structures when processing information streams, exhibiting superior performance, particularly in detecting small and dense targets. Furthermore, this module boasts excellent hardware adaptability and edge deployment performance, significantly reducing computational complexity while improving accuracy, thus providing a stable foundation for subsequent structural integration.

[0177] A deformable large-kernel target perception module is inserted after F3_out and F4_out to enhance the feature representation of small target regions; specifically:

[0178] To enhance the detection model's ability to perceive key target areas in complex industrial contexts, this embodiment introduces a deformable large kernel attention mechanism (D-LKA) at key locations in the cross-scale feature fusion network to enhance the spatial semantic representation and context modeling capabilities of the feature maps. Specifically, this mechanism is applied after the low-level fusion feature map F3 and after the mid-level fusion feature map F5.

[0179] In the deformable large kernel attention mechanism, for the input feature map, the receptive field of key spatial regions is first adaptively sampled through a deformable convolution module. A spatial offset matrix ΔP is generated from the input feature map through an offset prediction network, and this offset matrix is ​​introduced into the convolution sampling position to dynamically adjust the receptive field, thereby achieving fine alignment and information enhancement of the target region and improving the model's localization and boundary awareness capabilities.

[0180] The feature map processed by the deformable convolution module is then used to model contextual information through the following multi-scale large kernel convolution combination: First, a 5×5 deformable depth convolution is used to extract local structural features, and then a 7×7 deformable depth dilation convolution is used to extract region contours and long-distance dependencies, forming an attention mechanism equivalent to a 21×21 large receptive field.

[0181] All convolutional results are passed sequentially and finally fused and compressed using 1×1 convolutions to ensure compact feature representation and complete information. During the fusion process, the feature maps are weighted directly using element-wise multiplication to generate output feature maps with significant attention responses, which are then fed into subsequent detection heads for inference.

[0182] The feature maps enhanced by the deformable large kernel attention mechanism (F3_out and F4_out respectively) will replace the original F3 and F5 to construct the final multi-scale detection branch, effectively improving the response capability and discrimination accuracy of the target region.

[0183] A deformable large kernel attention mechanism is introduced at key locations in a cross-scale feature fusion network. (See [link]) Figure 10 The heatmap comparison before and after introducing the deformable large kernel attention mechanism is shown. Introducing the deformable large kernel attention mechanism gives the model a stronger receptive field and key region perception capability, effectively handling complex background interference and low-contrast targets in images. Compared to the baseline model without the deformable large kernel attention mechanism, the fusion model achieves significant improvements in performance metrics such as mAP50 and accuracy, especially demonstrating stronger detection advantages in small target detection scenarios.

[0184] The three-way fusion features F3_out, F4_out, and F5_out are fed into the multi-task decoupled prediction network in parallel. The classification branch outputs the class probability, and the regression branch outputs the center point, width, and height. After optimization by the EIOU loss function, the final detection box and class are obtained, and the class, location, and confidence recognition results are output.

[0185] Compared to the existing CIOU loss function, this invention replaces it with the more efficient EIOU loss function, specifically:

[0186] In existing detection networks, the bounding box regression task uses the CIOU loss function, which is calculated as follows:

[0187]

[0188] in:

[0189] IoU is the intersection-union ratio between the predicted bounding box and the ground truth bounding box;

[0190] ρ 2 (b,b gt () represents the squared Euclidean distance between the center point of the predicted bounding box and the center point of the ground truth bounding box;

[0191] c is the length of the diagonal of the smallest bounding rectangle;

[0192] v indicates the aspect ratio consistency item;

[0193] α is the balance coefficient.

[0194] This loss function suffers from problems such as coupling of optimization terms and limited convergence speed when considering center point distance and aspect ratio.

[0195] Therefore, this invention replaces CIOU loss with EIOU loss. In EIOU loss, the V and α terms from CIOU are no longer used; instead, explicit regression optimization is performed directly on the width and height difference. The calculation formula is as follows:

[0196]

[0197] in:

[0198] (ww gt ) 2 and (hh) gt ) 2 This is the squared error term between the predicted bounding box and the ground truth bounding box in the width and height directions;

[0199] w c and h c The image width and height are used to normalize scale differences.

[0200] Replace the CIOU loss calculation module used in the original regression loss function with the EIOU loss module. After the replacement, retrain the overall detection model and observe the changes in model convergence speed and bounding box localization accuracy.

[0201] The network model of this invention is trained using the augmented dataset to obtain a trained instrument target detection model. Specifically:

[0202] Input the images and corresponding labeled files from the dataset into the model training framework, and adjust the image size to 640×640.

[0203] Set training hyperparameters: initial learning rate of 0.01, momentum coefficient of 0.937, weight decay coefficient of 0.0005, set the number of training rounds to 150, and the batch size to 8; use the Stochastic Gradient Descent (SGD) optimization algorithm for model training to improve the model convergence effect;

[0204] Through continuous iterative training, finally obtain a YOLOv8-MCDE network model with good convergence and excellent performance, and save the trained weight file for subsequent inference and deployment.

[0205] Step 3: Deploy the trained object detection model on edge devices with limited computing resources for instrument object detection tasks. Specifically:

[0206] Export the weight file of the trained instrument object detection model to an inference format supported by edge devices;

[0207] Integrate an inference engine in the inspection robot platform, load the instrument object detection model, and configure the inference environment parameters to ensure that the model can run efficiently on embedded devices;

[0208] Real-time collect substation field instrument images through the front camera of the inspection robot, send them to the deployed instrument object detection model for inference, and output the detection results in real-time;

[0209] The detection results include the position coordinates, category information, and confidence of the instrument object, and the results can be used for subsequent display, storage, or linkage control by the inspection system; see some results in Figure 11 as shown.

[0210] The actual deployment test results show that the instrument object detection model can achieve high-precision and low-latency real-time instrument detection on edge devices, meeting the application requirements of industrial inspection robots for embedded deployment.

[0211] The following presents a comparison of the present invention with existing models for instrument detection.

[0212] To verify the performance of the lightweight instrument small object detection method in complex scenarios of the instrument object detection model proposed in the present invention, the YOLOv8s model, YOLO series models, and the model of the present invention are respectively used for performance comparison tests. To comprehensively compare the model performance, the present invention uses the following evaluation indicators: Precision (precision rate), Recall (recall rate), mAP (mean average precision), GFLOPs (computational volume), and Weight Size / MB (model weight size).

[0213] Precision measures how many of the detected positive samples are truly positive samples, and Recall measures the proportion of actual positive samples that are correctly detected. The specific calculation formulas are as follows:

[0214]

[0215] Where TP is the number of correctly detected positive samples, FP is the number of falsely detected positive samples, and FN is the number of missed positive samples.

[0216] mAP (mean Average Precision) is the average AP across all categories, calculated using the following formula:

[0217]

[0218] Where C is the total number of categories, AP c This represents the average accuracy for the C-class target.

[0219] GFLOPs are used to measure the computational complexity of a model and can reflect the model's computational resource requirements in actual deployment. The lower the GFLOPs value, the faster the model inference speed and the higher the deployment efficiency.

[0220] Weight Size indicates the storage size of the model weight file, which reflects the model's storage usage. The smaller the weight, the more suitable it is for deployment on edge devices.

[0221] By comprehensively analyzing the detection accuracy and computational efficiency of each model in complex background and small target detection scenarios using the above evaluation indicators, the effectiveness and engineering applicability of the method of this invention are verified.

[0222] The experimental environment parameters for this embodiment are shown in Table 1.

[0223] Table 1 Experimental Environment Parameters

[0224]

[0225] See the comparison chart of the results of this invention and existing YOLOv8s model testing instruments. Figure 12 As shown in Table 2, a performance comparison of the YOLOv8s model, the YOLO series models, and the model of this invention is presented.

[0226] Table 2 Performance Comparison of YOLOv8s Model, YOLO Series Models and the Model of this Invention

[0227]

[0228] As shown in Table 2, the YOLOv8s model outperforms the YOLOv5s, YOLOv10s, and YOLO11s models in key metrics such as mAP50-95, precision, and recall, indicating that YOLOv8s possesses superior detection capabilities and strong engineering application value. The instrument target detection model of this invention outperforms the YOLOv8s model in both precision and mAP50, with a 2.38 percentage point improvement in precision and a 1.27 percentage point improvement in mAP50. Simultaneously, the computational complexity of the instrument target detection model of this invention is reduced from 28.4G to 17.7G, and the model weight size is reduced from 22.5MB to 14.4MB, significantly improving the model's lightweight nature and deployment efficiency, making it particularly suitable for real-time detection applications on edge devices and inspection robot platforms.

[0229] The results in summary show that the instrument target detection model of the present invention effectively reduces computational complexity and model size while ensuring detection accuracy, and has excellent engineering application prospects and promotional value.

[0230] While the invention has been described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the invention. Therefore, it should be understood that many modifications can be made to the exemplary embodiments, and other arrangements can be designed without departing from the spirit and scope of the invention as defined by the appended claims. It should be understood that different dependent claims and features described herein can be combined in ways different from those described in the original claims. It is also understood that features described in conjunction with individual embodiments can be used in other described embodiments.

Claims

1. A lightweight instrument small target detection model for complex industrial scenarios, characterized in that, It includes a lightweight backbone network, a cross-scale feature fusion network, a deformable large-kernel target perception module, and a multi-task decoupled prediction network; Lightweight backbone network: used to extract multi-scale features from input images with limited computing resources; Cross-scale feature fusion network: connected to the lightweight backbone network, used to bidirectionally fuse the multi-scale features, retaining shallow spatial details and deep semantic information, and outputting fused features; Deformable large-kernel target perception module: Deployed at designated locations in the cross-scale feature fusion network to better capture feature information related to the target region, including: Global Context Extraction Unit: A large receptive field is constructed using depthwise separable convolution and dilated convolution to capture multi-scale contextual information; Adaptive Spatial Sampling Unit: Learns spatial offset through deformable convolution to dynamically adjust feature sampling position; Feature Enhancement Output Unit: Integrates global context and adaptive sampling results to highlight the target region and suppress background interference, thereby improving the feature representation capability of small targets; The multi-task decoupled prediction network performs classification, localization regression, and confidence prediction in parallel on the input image. The localization regression branch uses the decoupled width and height optimized EIOU loss function.

2. The lightweight instrument small target detection method for complex industrial scenarios according to claim 1, characterized in that, Lightweight backbone networks include: Standard convolutional layers are used to perform initial downsampling of the input image; Multiple inverted residual bottleneck modules, or Bneck modules, are stacked sequentially to form at least four stages to achieve multi-level downsampling; The Spatial Pyramid Pooling Module (SPPF) is used to integrate global context information; The lightweight backbone network outputs a shallow feature map P3 with a downsampling rate of 1 / 8, a mid-level feature map P4 with a downsampling rate of 1 / 16, and a deep feature map P5 with a downsampling rate of 1 / 32.

3. The lightweight instrument small target detection model for complex industrial scenarios according to claim 2, characterized in that, The standard convolutional layer of the lightweight backbone network is a 3×3 convolutional layer with 16 output channels and a stride of 2. The four stages of the Bneck modules stacked sequentially are: Phase 1: 2 Bneck modules. Each Bneck module in this phase contains a 3×3 depthwise separable convolution with 24 output channels. Second stage: 3 Bneck modules. Each Bneck module in this stage contains a 5×5 depthwise separable convolution with 40 output channels. The third stage consists of two Bneck modules, each containing a 5×5 depthwise separable convolution with 48 output channels. Fourth stage: 3 Bneck modules. The Bneck modules in this stage contain 5×5 depthwise separable convolutions with 96 output channels. Then, the number of channels is adjusted to the required dimension by the feature compression layer and then sent to the Spatial Pyramid Pooling Module (SPPF) for multi-scale pooling and global context integration. Feature compression layer: 1×1 convolution, the 1×1 convolution output channel 576; The output channel of the Spatial Pyramid Pooling Module (SPPF) is 1024.

4. The lightweight instrument small target detection model for complex industrial scenarios according to claim 2, characterized in that, The cross-scale feature fusion network includes two bidirectional fusion paths: First fusion path: The spatial resolution of the deep low-resolution feature map is gradually increased by upsampling units to align with the shallow high-resolution feature map, and then the feature maps are stitched together to form the first fusion path from top to bottom. The second fusion path: The spatial resolution of the shallow high-resolution feature map is gradually reduced by the downsampling unit to align with the deep low-resolution feature map, and then the feature maps are stitched together to form a bottom-up second fusion path. The cross-scale feature fusion network outputs fused features corresponding to the three detection scales P3, P4, and P5, respectively.

5. The lightweight instrument small target detection model for complex industrial scenarios according to claim 4, characterized in that, The first fusion path of the cross-scale feature fusion network includes two upsampling units, which are fused using a cross-stage local double convolutional layer C2f module. The first upsampling unit upsamples the deep feature map P5 by a factor of 2 to obtain a first upsampled feature with the same spatial size as the middle feature map P4. The first upsampled feature is then concatenated with the middle feature map P4 along the channel dimension and fed into the C2f module for fusion, generating a first fused feature F4. The second upsampling unit further upsamples F4 by a factor of 2 to obtain a second upsampled feature with the same spatial size as the shallow feature map P3. The second upsampled feature is then concatenated with the shallow feature map P3 along the channel dimension and fed into the C2f module to generate a second fused feature F3. The first fusion path is set with a first convolutional compression unit, which is used to perform 1×1 convolution on F3 and F4 in the first fusion path respectively to reduce the number of channels; The second fusion path of the cross-scale feature fusion network includes a downsampling unit, which uses the C2f module for fusion. A bottom-up reverse path is constructed using the downsampling unit, where the channel-compressed F3 is downsampled with a step size of 2. After downsampling, F3 is concatenated with the channel-compressed F4 and fed into the C2f module to generate the third fusion feature F5. The third fusion feature F5 is then downsampled with a step size of 2 and concatenated with the deep feature map P5, fed into the C2f module to generate the fourth fusion feature F6. The second fusion path includes a second convolutional compression unit, which performs a 1×1 convolution on the third fusion feature F5 to unify the channel dimensions.

6. The lightweight instrument small target detection model for complex industrial scenarios according to claim 5, characterized in that, The cross-scale feature fusion network has feature input and output ends set before and after the two fusion paths, respectively; At the feature input end, three 1×1 convolutions are used to receive shallow feature maps P3 (1 / 8 downsampling rate), mid-layer feature maps P4 (1 / 16 downsampling rate), and deep feature maps P5 (1 / 32 downsampling rate) from the lightweight backbone network, respectively. The output corresponds to the three final fused feature maps, each with a resolution of 1 / 8. 1 / 16 resolution and 1 / 32 resolution For use with decoupled prediction heads.

7. The lightweight instrument small target detection model for complex industrial scenarios according to claim 5, characterized in that, The deformable large-kernel target perception module is set at a designated node in the cross-scale feature fusion network, and the designated node is: In the first fusion path, F4 is the first feature node after upsampling and fusion with P3; as well as In the second fusion path, the second feature node is obtained by splicing and fusing F3 after downsampling with F4 after channel compression; The deformable large kernel target perception module is implemented using a deformable large kernel attention mechanism, which is applied to the fused feature maps of the first feature node and the second feature node, respectively.

8. The lightweight instrument small target detection model for complex industrial scenarios according to claim 7, characterized in that, The deformable large-kernel target perception module includes: A deformable convolution submodule is used to dynamically adjust the convolution sampling position based on the offset matrix; A multi-scale large-kernel convolutional submodule is used to extract contextual features from different receptive fields in parallel. The attention-weighted submodule is used to apply attention weights to the input feature map through element-wise multiplication to enhance the feature representation of small target regions.

9. A lightweight instrument small target detection method for complex industrial scenarios, the method is implemented based on any model with weights 1 to 7, characterized in that, The method includes the following steps: Step 1: Construct an image dataset containing small-sized instrument targets, and re-annotate and augment the images in the dataset; Step 2: Input the enhanced dataset into the object detection model for object recognition training, where: A lightweight backbone network receives the original image, downsamples it through a multi-level inverse residual bottleneck module Bneck, and outputs features P3, P4, and P5 at three scales: 1 / 8, 1 / 16, and 1 / 32, respectively. The cross-scale feature fusion network first passes P3, P4, and P5 through a 1×1 convolution to unify the channels, and then performs the following sequentially: Top-down approach: Upsampling - concatenation - fusion - convolutional compression to generate fused features F3 and F4; Bottom-up approach: downsampling - concatenation - fusion - convolutional compression to generate fused features F5 and F6; The final output is a three-way fused feature. , , ; Deformable large-core target perception module, in , Then, deformable large-kernel target perception modules are inserted to enhance the feature representation of small target regions; Multi-task decoupled prediction network, three-way feature fusion , , The data is fed into a multi-task decoupled prediction network in parallel. The classification branch outputs the class probability, and the regression branch outputs the center point, width, and height. After optimization by the EIOU loss function, the final detection box and class are obtained, and the class, location, and confidence recognition results are output. Step 3: Deploy the trained target detection model on an edge device with limited computing resources to perform instrument target detection tasks.

10. The lightweight instrument small target detection method for complex industrial scenarios according to claim 9, characterized in that, The edge device is an inspection robot.

Images