Noodle adhesion recognition method based on improved YOLOv10m lightweight small target enhancement detection model

Abstract

The application discloses a kind of based on improved YOLOv10m light weight small target enhancement detection model noodle adhesion identification method, belong to food industrialization intelligent production and machine vision target detection cross technical field. Including: the original image of acquisition noodle industrialization cutting link;Establish noodle adhesion state detection model and training;Using the model detects noodle adhesion state.The noodle adhesion state detection model of the application, with YOLOv10m as baseline model, adopts light weight improved NAD-MobileViT main network, to adapt to the extraction demand of noodle slender structure characteristics;Introduce CEECA module in Neck feature fusion layer, realize the channel-space joint reinforcement of core feature and redundancy information suppression;In Head detection head part, add new high-resolution small target detection head.It can realize the real-time accurate monitoring of noodle edge form after cutting, can effectively identify interlaced state.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of intelligent production in the food industry and machine vision target detection, specifically to the detection of quality defects in the industrial cutting process of noodles. Based on an improved YOLOv10m model, it achieves real-time identification and accurate quantification of noodle sticking and width deviation, providing data support for the dynamic control of noodle production parameters. This allows for timely adjustment of process and raw material parameters during noodle production to prevent product sticking and realize intelligent noodle production. Background Technology

[0002] In the industrialized noodle production system, the processing quality of the cutting stage directly determines the product quality and the continuity of subsequent drying, cutting, and packaging processes. Inter-noodle adhesion is a frequent quality problem in this stage. Currently, noodle cutting equipment mostly operates with fixed mechanical parameters, making it highly susceptible to edge adhesion issues due to fluctuations in dough moisture content, blade wear, and uneven conveying speed. Traditional detection methods rely primarily on manual sampling, which suffers from low efficiency, strong subjectivity, and the inability to intervene in real time. If problems are not detected early in production, it can easily lead to batch product scrapping and significantly increase production losses. Therefore, developing a visual inspection method that can detect the morphology of noodle cutting edges in real time and intelligently identify adhesion is of significant engineering importance for improving the automation and intelligence level of noodle production.

[0003] In recent years, the YOLO series of single-stage object detection algorithms have become the mainstream technical solution in the field of industrial vision inspection due to their end-to-end training paradigm, efficient inference speed and excellent detection accuracy. With the iterative upgrades of new generation models such as YOLOv10, YOLOv11 and YOLOv13, their performance in lightweight design, multi-scale feature fusion and small object detection has been greatly improved, and they have been widely used in defect detection scenarios throughout the entire food production process. In existing research, scholars have conducted many cutting-edge explorations in food and slender structure detection based on the new generation YOLO model. For example, Li et al. designed a lightweight feature enhancement module based on YOLOv11 to achieve real-time online detection of micro-defects in grain and oil particles, with detection accuracy and inference speed improved by more than 15% compared with traditional models; Wang et al. introduced the Transformer global modeling structure on the basis of YOLOv10, which effectively improved the robustness of detecting minor damage on the surface of fruits and vegetables in complex industrial backgrounds; Zhang et al. applied the lightweight YOLOv13 model to the fracture detection task in noodle production, and achieved efficient deployment on embedded devices through model pruning and operator optimization, with a detection frame rate of more than 30 FPS. In the field of slender structure detection, cutting-edge research has also achieved technological breakthroughs by relying on the new generation of YOLO models. For example, Chen et al. combined YOLOv13 with the lightweight SwingTransformer to build a steel wire arrangement anomaly detection model, which solved the problem of cross-region feature modeling of dense slender structures. Hu et al. added a super-resolution detection branch based on YOLOv11, which enabled accurate identification of spacing deviation of microscale strip food, and improved the detection AP value of sub-millimeter targets to 89.2%.

[0004] While the aforementioned research has achieved some success in the fields of strip structure and industrial inspection, it still has significant shortcomings in the specific scenario of noodle anti-adhesion detection. First, noodles are arranged in a highly regular parallel and slender structure, and adhesion is mostly manifested as millimeter-level edge adhesion. Uneven industrial lighting and background texture interference can easily lead to blurred noodle outlines, placing extremely high demands on the model's fine-scale feature extraction capabilities. Second, the backbone network feature extraction unit parameters of the original detection model are fixed, making it difficult to adapt to the dynamic changes in the slender structure of noodles, and easily leading to problems such as loss of edge details and background noise interference. Third, the acquired noodle detection images contain a large amount of redundant information, which can easily obscure core features, and traditional models lack targeted feature attention mechanisms, making it difficult to achieve core feature enhancement and redundant feature suppression. Finally, the embedded deployment and high-speed detection requirements of industrial production require the model to ensure fine detection accuracy while also taking into account lightweight design and efficient inference performance. Summary of the Invention

[0005] To address at least one of the aforementioned problems, the present invention aims to provide an intelligent production method for preventing noodle sticking based on an improved YOLOv10m lightweight small target enhanced detection model.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A method for identifying noodle adhesion based on an improved YOLOv10m lightweight small target enhancement detection model includes the following steps: S1. Collect original images of the industrial cutting process of noodles and mark the parts of the noodles that are stuck together in the images; S2. Establish a noodle adhesion state detection model based on YOLOv10m as the baseline model. The noodle adhesion state detection model includes a backbone network, a feature fusion layer, and a detection head. The backbone network is MobileViT, and the MV2 module of MobileViT introduces a neural architecture search and dynamic parameter adjustment mechanism (NAD) to form an NAD-MV2 module, which is adapted to the feature extraction requirements of the slender structure of noodles. The feature fusion layer introduces a coordinate-enhanced efficient channel attention (CEECA) module to achieve joint channel-space enhancement of core features and suppression of redundant information. A high-resolution small target detection head is added to the detection head to improve the detection accuracy of millimeter-level adhesion and size deviation. S3. Use the images labeled with the noodle adhesion status in step S1 to train the noodle adhesion status detection model in step S2. S4. Use a camera to capture on-site images of the noodles and input them into the trained noodle adhesion detection model to detect the adhesion status of the noodles.

[0007] In a specific embodiment of the present invention, in step S2, the NAD-MV2 module searches for the optimal expansion rate (t∈{2,4,6}) and output channel number (C∈{64,128,256}) for the feature extraction level it is in, thereby achieving low expansion rate compression of shallow high-resolution features and preserving fine-scale details such as noodle edges and width to the greatest extent. The high expansion rate of deep low-resolution features is enhanced to strengthen the global semantic features of noodle arrangement. At the same time, the NAD-MV2 module sets two optional convolution kernels, 3×3 and 5×5, and adopts a dynamic convolution kernel selection strategy to improve the adaptability of the backbone network to the geometric features of slender structural targets such as noodles, effectively adapting to the morphological characteristics of high-density parallel arrangement of noodles.

[0008] In one specific embodiment of the present invention, two coordinate-enhanced high-efficiency channel attention modules are provided, both located in the top-down upsampling path of the feature fusion layer. The first coordinate-enhanced high-efficiency channel attention module is embedded in the PSA module, and its output participates in the bottom-up downsampling fusion of the feature fusion layer. The second coordinate-enhanced high-efficiency channel attention module is located before the third splicing of the top-down upsampling path.

[0009] As a specific embodiment of the present invention, the specific operation steps of the CEECA module are as follows: S221. Perform one-dimensional global pooling operations, namely horizontal global average pooling and vertical global average pooling, on the input feature map to obtain two independent orientation-aware feature maps, which completely preserve the spatial structure information of the noodles along the width and length directions. S222. Perform batch normalization and nonlinear activation processing on the two directional sensing feature maps respectively to fuse spatial location information and channel features; S223. Cross-channel interaction is achieved through ECA one-dimensional adaptive convolution, which autonomously learns the dependencies between channels and generates corresponding channel weights; S224. The weighted position-aware features are fused with the original input feature map using a weighted method, and the output is obtained via a channel. Optimized feature maps with spatial joint attention enhancement.

[0010] In one specific embodiment of the present invention, the newly added P2 branch is a 160×160 high-resolution small target detection branch.

[0011] Compared with existing technologies, this invention has the following advantages: This invention proposes a noodle adhesion recognition method, using YOLOv10m as the baseline model, and achieves a comprehensive improvement in detection performance through three targeted structural optimizations: A lightweight improved NAD-MobileViT backbone network is designed to replace the original backbone, adapting to the requirements of extracting the slender structural features of noodles; a newly proposed CEECA module is introduced into the Neck feature fusion layer to achieve channel-space joint enhancement of core features and suppression of redundant information; a high-resolution small target detection head is added to the Head detection head to improve the detection accuracy of millimeter-level adhesion and size deviation. The proposed improved YOLOv10m model enables accurate calculation of the real-time detection of the edge morphology of noodles after cutting, effectively identifying inter-strip adhesion and providing data support for real-time control of production parameters. It also has the advantages of high detection accuracy, fast inference speed, and adaptability to industrial embedded deployment, providing a feasible solution for intelligent control of noodle production quality. Attached Figure Description

[0012] Figure 1This is a schematic diagram of the overall network structure of the improved YOLOv10m detection model in an embodiment of the present invention; Figure 2 This is a schematic diagram of the NAD-MV2Block module in an embodiment of the present invention; Figure 3 This is a schematic diagram of the CEECA coordinate-enhanced high-efficiency channel attention module in an embodiment of the present invention. Detailed Implementation

[0013] To more clearly illustrate the present invention, specific embodiments are described below. Those skilled in the art should understand that the following description is illustrative rather than restrictive and should not be construed as limiting the scope of protection of the present invention.

[0014] The noodle adhesion recognition method based on the improved YOLOv10m lightweight small target enhancement detection model in this embodiment includes the following steps: S1. Collect original images of the industrial cutting process of noodles and mark the parts of the noodles that are stuck together in the images; In this step, image acquisition uses a global shutter industrial area array camera paired with a ring-shaped shadowless light source, mounted 40cm directly above the noodle cutting machine outlet for vertical shooting. The resolution is set to 4096×2160, which can clearly show 0.1mm-level adhesion and edge bonding. 15,000 effective images covering normal, light / severe adhesion conditions are collected and divided into training, validation, and test sets in a 7:2:1 ratio. The LabelImg tool is used to finely annotate the adhesion areas in the images with tight rectangular boxes, and output YOLO format label files.

[0015] S2. Establish a noodle adhesion state detection model. The model uses YOLOv10m as the baseline, and while retaining its end-to-end detection framework, PSA global feature modeling advantages, and lightweight core design, targeted improvements are made to the Backbone network, Neck feature fusion layer, and Head detection head. The overall structure is as follows: Figure 1 As shown, the backbone network is MobileViT, and the MV2 module of MobileViT incorporates Neural Architecture Search (NAD) and dynamic parameter adjustment mechanisms to form the NAD-MV2 module, adapting to the feature extraction requirements of the slender noodle structure. The feature fusion layer introduces a coordinate-enhanced efficient channel attention (CEECA) module to achieve joint channel-space enhancement and redundancy suppression of core features. A new high-resolution small target detection head is added to the detection head to improve the detection accuracy of millimeter-level adhesion and size deviation. The following detailed description of each module of the model, with reference to the attached figures, is provided: like Figure 2As shown, the NAD-MV2 module integrates Neural Architecture Search (NAD) with a dynamic parameter adjustment mechanism. Its operating mechanism involves hierarchical search to match the optimal expansion rate (t∈{2,4,6}) and number of output channels (C∈{64,128,256}) for the feature extraction level at its current location. Specifically, shallow high-resolution features (levels 1 and 2) are compressed with a low expansion rate (t=2), and the output channels are set to 64 to maximize the preservation of fine-scale details such as noodle edges and width. Deep low-resolution features (levels 3 and 4) are augmented with a high expansion rate. The output channel is set to 256 (t=6) to enhance the global semantic features of the noodle arrangement. At the same time, the NAD-MV2 module sets two optional convolution kernels, 3×3 and 5×5, and adopts an adaptive receptive field dynamic selection strategy: based on the scale of the current feature map and the directional distribution of the noodle's slender structure, it automatically selects a 3×3 convolution kernel to focus on local edge details, or switches to a 5×5 convolution kernel to expand the receptive field to adapt to the global geometric features of the long strip target, thereby improving the adaptability of the backbone network to the geometric features of slender structural targets such as noodles, and effectively adapting to the morphological characteristics of the high-density parallel arrangement of noodles.

[0016] like Figure 1 As shown, two coordinate-enhanced high-efficiency channel attention modules are set, both located in the top-down upsampling path of the feature fusion layer. The first coordinate-enhanced high-efficiency channel attention module is embedded in the PSA module, and its output participates in the bottom-up downsampling fusion of the feature fusion layer. The second coordinate-enhanced high-efficiency channel attention module is located before the third concatenation of the top-down upsampling path.

[0017] like Figure 3 As shown, the CEECA module first performs one-dimensional global pooling operations on the input feature map through horizontal global average pooling (XAvgPool) and vertical global average pooling (YAvgPool), aggregating the features into two independent orientation-aware feature maps, preserving the spatial structure information of the noodles along the width and length directions. Then, batch normalization and non-linear activation are used to encode the aware features, fusing spatial location information with channel features. Next, one-dimensional adaptive convolution is used to achieve cross-channel interaction, autonomously determining the dependencies between channels and generating channel weights. Finally, the weighted location-aware features are weighted and fused with the original feature map, enabling the model to accurately focus on key areas such as the noodle edges.

[0018] S3. Use the images labeled with the noodle adhesion status in step S1 to train the noodle adhesion status detection model in step S2. S4. Use a camera to capture on-site images of the noodles and input them into the trained noodle adhesion detection model to detect the adhesion status of the noodles.

[0019] The model trained in step S3 of this embodiment can achieve real-time online detection of noodle cutting quality. When adhesion defects are detected, the information is transmitted to the intelligent control system, which immediately issues an early warning and feeds back the detection data to the production line control system. This assists operators in dynamically adjusting production parameters such as cutting tool spacing, conveying speed, and dough moisture content, thereby achieving closed-loop control of production quality.

[0020] To verify the performance of the method of the present invention, a comparative experiment was set up in this embodiment. The baseline model YOLOv10m model was used as the comparison method. Training and testing were carried out under the same dataset and experimental environment, and the mean accuracy mAP@0.5 was selected as the core evaluation index.

[0021] Ablation experiments showed that replacing the NAD-MobileViT backbone, adding the CEECA module, and increasing the small target detection head improved the mean accuracy (mAP@0.5) by 2.2%, 3.6%, and 0.8% respectively compared to the baseline model, indicating that the three proposed modules independently and positively improved the performance of the baseline model. When the three modules were used simultaneously, i.e., when the improved YOLOv10m model was compared with the baseline YOLOv10m model, the mean accuracy (mAP@0.5) improved by 6.2%, reaching an excellent index of 96.7%, indicating that the three proposed modules were innovative without conflict and greatly optimized the overall model.

[0022] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for identifying noodle adhesion based on an improved YOLOv10m lightweight small target enhancement detection model, characterized in that, Includes the following steps: S1. Collect original images of the industrial cutting process of noodles and mark the parts of the noodles that are stuck together in the images; S2. Establish a noodle adhesion state detection model based on YOLOv10m as the baseline model. The noodle adhesion state detection model includes a backbone network, a feature fusion layer, and a detection head. The backbone network is MobileViT, and a neural architecture search and dynamic parameter adjustment mechanism is introduced into the MV2 module of MobileViT to form a NAD-MV2 module to adapt to the feature extraction requirements of the slender structure of noodles. A coordinate-enhanced high-efficiency channel attention module is introduced into the feature fusion layer to achieve channel-space joint enhancement of core features and suppression of redundant information. A high-resolution small target detection head is added to the detection head to improve the detection accuracy of millimeter-level adhesion and size deviation. S3. Use the images labeled with the noodle adhesion status in step S1 to train the noodle adhesion status detection model in step S2. S4. Use a camera to capture on-site images of the noodles and input them into the trained noodle adhesion detection model to detect the adhesion status of the noodles.

2. The noodle adhesion recognition method based on the improved YOLOv10m lightweight small target enhancement detection model according to claim 1, characterized in that, The NAD-MV2 module searches to match the optimal expansion rate and number of output channels for its current feature extraction level. At the same time, the NAD-MV2 module sets two optional convolution kernels, 3×3 and 5×5, and adopts a dynamic convolution kernel selection strategy.

3. The noodle adhesion recognition method based on the improved YOLOv10m lightweight small target enhancement detection model according to claim 1, characterized in that, Two coordinate-enhanced high-efficiency channel attention modules are configured, both located in the top-down upsampling path of the feature fusion layer. The first coordinate-enhanced high-efficiency channel attention module is embedded in the PSA module, and its output participates in the bottom-up downsampling fusion of the feature fusion layer. The second coordinate-enhanced high-efficiency channel attention module is located before the third concatenation of the top-down upsampling path.

4. The noodle adhesion recognition method based on the improved YOLOv10m lightweight small target enhancement detection model according to claim 1, characterized in that, The specific operation steps of the coordinate-enhanced high-efficiency channel attention module are as follows: S221. Perform one-dimensional global pooling operations, namely horizontal global average pooling and vertical global average pooling, on the input feature map to obtain two independent orientation-aware feature maps, which completely preserve the spatial structure information of the noodles along the width and length directions. S222. Perform batch normalization and nonlinear activation processing on the two directional sensing feature maps respectively to fuse spatial location information and channel features; S223. Cross-channel interaction is achieved through ECA one-dimensional adaptive convolution, which autonomously learns the dependencies between channels and generates corresponding channel weights; S224. The weighted position-aware features are fused with the original input feature map using a weighted method, and the output is obtained via a channel. Optimized feature maps with spatial joint attention enhancement.

5. The noodle adhesion recognition method based on the improved YOLOv10m lightweight small target enhancement detection model according to claim 1, characterized in that, The newly added high-resolution small target detection head is a 160×160 high-resolution small target detection branch.

Images