Lightweight target detection method and system for abnormal rice detection

Abstract

The invention discloses a lightweight target detection method and system for abnormal rice detection. Belongs to the technical field of neural network target detection and particularly relates to the technical field of abnormal rice detection. The technical problem that an existing neural network technology is low in abnormal rice detection precision and efficiency is solved. The method comprises the following steps: data set construction: acquiring different types of abnormal rice pictures, and carrying out category labeling and division of a training set, a verification set and a test set; model construction: on the basis of the YOLOv11n model, combining rice morphological characteristics, improving the YOLOv11n model, and constructing a model suitable for abnormal rice detection; model training: adopting the constructed data set to train a model suitable for abnormal rice detection, and adjusting model parameters until the model meets detection requirements; and carrying out abnormal rice detection by adopting the trained model suitable for abnormal rice detection.

Description

Technical Field

[0001] The invention belongs to the technical field of neural network target detection, and in particular relates to the technical field of abnormal rice detection. Background Art

[0002] As the global population continues to grow, food security and food quality issues are receiving increasing attention. As one of the most important foods in the world, rice is widely used in the daily diet of various countries, and its quality is directly related to people's health and quality of life. Rice is easily affected by various factors during production, processing and transportation, and may be contaminated by foreign matter, moldy, broken, etc. These abnormal rice not only affect product quality, but also may pose a threat to consumer health. Therefore, how to efficiently and accurately detect and screen out abnormal rice in large-scale production has become an important challenge in current food quality control.

[0003] Traditional rice quality detection methods mainly rely on manual screening, which is not only labor-intensive and inefficient, but also easily affected by human factors, and it is difficult to meet the requirements of modern production lines for detection speed and accuracy. In order to improve the automation level and accuracy of rice detection, computer vision and deep learning technologies have gradually been introduced into the field of food quality detection. Object detection algorithms, especially object detection models based on deep learning, have achieved remarkable results in image recognition, object detection and other fields, but their application in the food industry still faces some specific challenges, including the diversity of different varieties of rice, the appearance differences of abnormal rice, and the high efficiency requirements of the model in real-time detection.

[0004] In addition, the existing target detection model cannot accurately adapt to the sensitivity of abnormal conditions such as rice surface defects, foreign matter contamination and mildew, as well as the high amount of calculation during detection. As a result, the existing neural network technology has low detection accuracy and efficiency for abnormal rice. Summary of the invention

[0005] In order to solve the technical problem that the existing neural network technology has low detection accuracy and efficiency for abnormal rice, the present invention provides a lightweight target detection method for abnormal rice detection, the method comprising the following steps: S1. Dataset construction: Collect different types of abnormal rice images, label them, and divide them into training sets, validation sets, and test sets; S2. Model construction: Based on the YOLOv11n model, the YOLOv11n model is improved in combination with the morphological characteristics of rice to build a model suitable for abnormal rice detection; The improvements to the YOLOv11n model are as follows: Reduce the number of model parameters and computational complexity; Introducing the attention mechanism; Optimize information fusion architecture; S3, model training: using the constructed data set to train the model suitable for abnormal rice detection, and adjusting the model parameters until the model meets the detection requirements; S4. Use the trained model suitable for abnormal rice detection to perform abnormal rice detection.

[0006] Furthermore, the different types of abnormal rice include four types, namely broken rice, contaminated and discolored rice, intact edible rice, and moldy rice.

[0007] Furthermore, the method of reducing the number of parameters and the amount of calculation of the model is specifically as follows: replacing the fifth convolution module starting from the input in the YOLOv11n model backbone network with a deep convolution module.

[0008] Furthermore, the introduction of the attention mechanism is specifically as follows: adding a SimAM module after the C2PSA module in the YOLOv11n model backbone network.

[0009] Furthermore, the optimized information fusion architecture is specifically as follows: a BiFPN architecture is introduced into the neck part of the YOLOv11n model, and an information fusion method of the neck part of the YOLOv11n model is changed.

[0010] Furthermore, the information fusion method of the neck part of the YOLOv11n model is changed specifically as follows: S1, connect the first C3K2 module of the YOLOv11n model trunk from the input to the neck part; S2, add a convolution module between the modules connecting the trunk and the neck of the YOLOv11n model; S3, replace the Concat module in the neck part with the Fusion module, and perform weighted fusion of features at different levels through learnable weights; S4. Change the one-way information flow mode of the neck part of the YOLOv11n model into a two-way cross-scale interaction mode.

[0011] Furthermore, the information flow of the neck part of the model suitable for abnormal rice detection is divided into four paths: The modules passed through in the first path are convolution module, fusion module and C3K2 module respectively; The second path is divided into two branches after the convolution module; the first branch passes through the Fusion module in the first path and the C3K2 module in the first path in sequence; the second branch passes through the Fusion module, the C3K2 module, the Fusion module in the first path, and the C3K2 module in the first path in sequence; The third path is divided into two branches after the convolution module; the first branch passes through the Fusion module, the C3K2 module, the upsampling module, the Fusion module in the second path, the C3K2 module in the second path, the Fusion module in the first path, and the C3K2 module in the first path; the second branch passes through the Fusion module, the C3K2 module, the convolution module, and then enters the Fusion module in the fourth path; The fourth path is divided into two branches after the convolution module; the first branch passes through the upsampling module and then enters the Fusion module of the first branch in the third path; the second branch passes through the Fusion module and the C3K2 module in turn; The C3K2 module in the first path outputs information into the head network. The C3K2 module in the first branch also inputs information into the Fusion module in the third path through a convolution module. The output information of the C3K2 module in the second branch of the third path enters the head network; The C3K2 module output information of the second branch in the fourth path enters the head network.

[0012] The present invention also provides a lightweight target detection system for abnormal rice detection, the system comprising the following modules: Module for constructing data sets: collect different types of abnormal rice images, label them, and divide them into training sets, validation sets, and test sets; Model building module: Based on the YOLOv11n model, the YOLOv11n model is improved in combination with the morphological characteristics of rice to build a model suitable for abnormal rice detection; The improvements to the YOLOv11n model are as follows: Reduce the number of model parameters and computational complexity; Introducing the attention mechanism; Optimize information fusion architecture; Model training module: Use the constructed data set to train the model suitable for abnormal rice detection and adjust the model parameters until the model meets the detection requirements; A module for abnormal rice detection using a trained model suitable for abnormal rice detection.

[0013] The beneficial effects of the method of the present invention are: By optimizing the YOLOv11n target detection architecture, the deep convolution module, the SimAM parameter-free attention mechanism, and the efficient BiFPN architecture are innovatively introduced in the neck adaptability, and the information processing mode of the neck is changed again, so that the model can significantly reduce the computational cost and parameter amount while ensuring high detection accuracy. The experimental results show that the method of the present invention has achieved remarkable results in the abnormal rice detection task, and the mAP50 is improved by 5.27% compared with the original model, and its parameter amount is reduced by 36.86% (i.e., 0.952M) compared with the original model, showing high detection accuracy and low computational overhead. The lightweight design of the present invention enables the model to run efficiently under limited equipment resources, and has good practicality and scalability. Through the streamlined network structure, combined with innovative algorithm optimization methods, the computational complexity of the model is effectively reduced, and the abnormal rice detection task can be quickly performed, which is particularly important for the quality inspection system in large-scale production lines. In addition, the high precision and low computational cost of the method of the present invention enable it to be widely used in various production environments, especially in automated production processes, where it can replace traditional manual screening methods, improve production efficiency, reduce human interference factors, and further ensure food safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 Flow chart of a lightweight target detection method for abnormal rice detection in an embodiment of the present invention; Figure 2 Schematic diagram of a rice image acquisition device in an embodiment of the present invention; Figure 3 This is a schematic diagram of labels corresponding to different types of rice in an embodiment of the present invention; Figure 4 : is a structural diagram of the Rice-YOLO-AD model in an embodiment of the present invention; Figure 5 This is a simplified diagram of the YOLOv11n model structure in an embodiment of the present invention; Figure 6 : is a simplified structural diagram of the Rice-YOLO-AD model in an embodiment of the present invention; Figure 7 Schematic diagram of the structure of ordinary convolution and depth convolution in an embodiment of the present invention; Figure 8 Schematic diagram of the structure of the SimAM attention mechanism in an embodiment of the present invention; Fig. 9 A schematic diagram of the structure of BiFPN in an embodiment of the present invention; Fig.10 is a convergence curve of the Rice-YOLO-AD model during the training process in an embodiment of the present invention; Fig.11: is a comparison chart of the mAP50 curves of the Rice-YOLO-AD model in the embodiment of the present invention and other classic models; Fig.12 Schematic diagram of placing SimAM in five different positions in the YOLOv11 backbone in an embodiment of the present invention; Fig.13 This is a comparison chart of the detection results of complete edible rice by the YOLOv11n model and the Rice-YOLO-AD model in an embodiment of the present invention; Fig.14 This is a comparison chart of the broken rice detection results of the YOLOv11n model and the Rice-YOLO-AD model in an embodiment of the present invention; Fig.15 This is a comparison chart of the moldy rice detection results of the YOLOv11n model and the Rice-YOLO-AD model in an embodiment of the present invention. DETAILED DESCRIPTION

[0015] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the protection scope of the present invention.

[0016] Embodiment 1, This embodiment provides a lightweight target detection method for abnormal rice detection. Figure 1 As shown, the method comprises the following steps: S1. Dataset construction: Collect different types of abnormal rice images, label them, and divide them into training sets, validation sets, and test sets; S2. Model construction: Based on the YOLOv11n model, the YOLOv11n model is improved in combination with the morphological characteristics of rice to construct the Rice-YOLO-AD model suitable for abnormal rice detection; The improvements to the YOLOv11n model are as follows: Reduce the number of model parameters and computational complexity; Introducing the attention mechanism; Optimize information fusion architecture; S3, model training: Use the constructed data set to train Rice-YOLO-AD and adjust the model parameters until the model meets the detection requirements; S4. Use the trained Rice-YOLO-AD to detect abnormal rice.

[0017] As an efficient target detection algorithm, YOLO (You Only Look Once) is widely used in the field of object recognition due to its advantages of fast speed and high accuracy. Nevertheless, YOLO still has certain limitations when dealing with food detection, such as the balance between processing speed and detection accuracy, especially when facing large-scale data sets. Therefore, in response to the special needs of rice detection, this embodiment proposes an improved YOLOv11n model, Rice-YOLO-AD (Rice-based You Only Look Once Anomaly Detection), which aims to solve the problem of efficiency and accuracy of rice surface anomaly detection. Compared with previous target detection algorithms, YOLOv11 introduces a more efficient architecture, including efficient attention mechanisms such as C3K2, SPPF and C2PSA. YOLOv11 aims to enhance small object detection and improve accuracy while maintaining YOLO's consistent real-time inference speed. As a lightweight model in the YOLOv11 series, YOLOv11n has a lower number of parameters and computational complexity while maintaining high accuracy, making it easier to deploy in resource-constrained environments.

[0018] Rice-YOLO-AD combines the real-time advantages of the YOLOv11 framework and optimizes it for the characteristics of rice. By adding an anomaly detection module, it improves the sensitivity to abnormal conditions such as rice surface defects, foreign matter contamination, and mildew. Compared with traditional image processing methods, Rice-YOLO-AD significantly improves computing efficiency while ensuring high detection accuracy, adapting to the needs of industrial production.

[0019] Embodiment 2, This embodiment further limits the embodiment 1 and further limits step S1.

[0020] Rice image acquisition device Figure 2 As shown, it consists of an image capturing device, a lighting device and a black light-absorbing plate. The entire shooting environment is carried out in a dark room. The rice sample category selected in this embodiment is Wuchang long-grain fragrant. The rice is randomly sprinkled on the black light-absorbing cloth without overlapping or adhering to the rice, and its RGB image is obtained by the shooting device. In this study, the shooting device used is a mobile phone Honor Magic6, which is fixed on the top bracket for vertical shooting, and the obtained image is annotated using the Make Sense online annotation website.

[0021] In this embodiment, rice is classified into the following four categories.

[0022] Broken rice (Category 0): Broken rice is usually caused by external forces during harvesting, processing or transportation. Broken rice not only affects the taste, but also may bring some potential food safety issues.

[0023] Contaminated and discolored rice (category 1): Discolored rice often changes color due to contamination, improper storage conditions, or attack by microorganisms such as mold and bacteria. This type of rice may contain harmful substances and requires special attention.

[0024] Whole, edible rice (Category 2): This category of rice represents good product quality, as it is not affected by breakage, contamination or mold.

[0025] Moldy rice (category 3): Moldy rice is usually caused by mildew due to storage conditions that are damp, warm or poorly ventilated. Moldy rice not only affects the safety of rice for consumption, but may also cause mycotoxins that are harmful to human health.

[0026] This detailed classification can help the model improve the recognition accuracy of each type of abnormal rice, avoid misclassifying abnormal rice of similar appearance as other categories, and thus improve the accuracy and robustness of detection. At the same time, this classification also provides more detailed data for model design, which helps to better optimize and debug the model. Dividing rice into four categories: broken, contaminated and discolored, edible, and moldy not only meets the needs of actual applications, but also helps to improve the accuracy and pertinence of the target detection algorithm. This classification method enables the model to identify and process different types of abnormal rice separately in complex detection tasks, thereby achieving more accurate and efficient quality control.

[0027] like Figure 3 As shown, label 0 represents broken rice, label 1 represents contaminated and discolored rice, label 2 represents intact and edible rice, and label 3 represents moldy rice.

[0028] (a) is the effect of all normal rice, (b) is the effect of broken rice and normal rice mixed together, (c) is the effect of all four types of rice, and (d) is the effect of the four types of rice mixed particularly densely.

[0029] The dataset has a total of 64 images, which are divided into training set, validation set and test set in a ratio of 4:1:1. In the rice dataset, the features between samples are highly consistent and representative. In addition, there are a large number of rice in each picture and the distribution is relatively dense, which enables each sample to fully demonstrate the key features of rice. Therefore, even if the number of pictures is relatively limited, the annotated data can still contain rich and effective information. This effective information can help the model effectively identify abnormal rice, thereby achieving accurate detection of rice quality. By making full use of these features and information, the efficiency of the model and the accuracy of detection can still be guaranteed even when the amount of data is not large. At the same time, it adapts to application scenarios in the real world, avoids the difficulty of collecting a large number of annotated samples, and reduces dependence on large-scale annotated data.

[0030] Embodiment 3, This embodiment further limits the embodiment 1 and further illustrates step S2.

[0031] like Figure 4 The structure diagram of the Rice-YOLO-AD model in this embodiment is shown. Due to limited space, each module is given a short name, and the full English name of each module is given here: FFN:Feed-ForwardNetwork,PSA:PyramidSpatialAttentionMechanism,C2PSA:CompoundTwo-PathSpatialAttention,SPPF:SpatialPyramidPooling-Fixed,C3K:C3KBlock,C3K2:C3KBlockAppliedTwice,SimAM:ASimple,Param eter-FreeAttentionModuleforConvolutionalNeuralNetworks,BatchNorm2d:2-DimensionalBatchNormalizationLayer,SiLU:SiLUActivationFunction,k:SizeoftheConvolutionalKernel,s:SizeofStep,p:SizeofPadding.

[0032] This embodiment is based on the YOLOv11n model to improve its ability to identify and detect rice. Figure 5 This is a simplified diagram of the YOLOv11n model structure. To make the model more efficient and improve recognition accuracy, the main improvement strategies can be summarized as follows: First, a normal convolution (Conv) module in the backbone network of the model was replaced with a deep convolution (DWConv) module. As an efficient convolution operation mode, deep convolution can significantly reduce the number of model parameters and computational complexity compared to traditional ordinary convolution. The main purpose of this improvement is to improve the operational efficiency of the model when processing rice recognition tasks. By reducing computational complexity, DWConv can make the training and reasoning of the model on the dataset more efficient, thereby speeding up the response speed and saving computing resources in the rice detection process. However, although DWConv can improve the computational efficiency of the model to a certain extent, it may also cause a decrease in the recognition accuracy of the model.

[0033] Ordinary convolution and deep convolution are two common convolution methods in convolutional neural networks. They differ in structure, calculation method and application scenarios, and each has its own unique advantages and disadvantages. The schematic diagram is as follows Figure 7 As shown in the figure. Ordinary convolution can flexibly adjust the size, step size, padding and other parameters of the convolution kernel according to task requirements. By adjusting the number and parameters of the convolution kernel, the complexity of the model and the feature extraction capability can be controlled. Deep convolution has extremely low parameter and computational complexity. It greatly reduces the number of parameters and computational complexity of the model by performing independent convolution operations on each input channel, so it is more suitable for lightweight models. Since each channel performs convolution operations independently, deep convolution can better maintain the spatial information of the input data.

[0034] During model training and inference, the size of the parameters directly determines the amount of computation required. Reducing the number of parameters can speed up training and inference, improve overall computational efficiency, and make the model's inference capabilities for rice faster and more suitable for use in resource-constrained environments. Deep convolution performs independent convolution operations on each input channel, while ordinary convolution performs convolution operations on the entire input feature map. The number of convolution kernels in deep convolution is the same as the number of input channels, and each convolution kernel is only responsible for the convolution operation of one channel, while the number of convolution kernels in ordinary convolution is more than the number of input channels, and each convolution kernel can process data from multiple channels. Assume that the number of input and output channels are respectively and , is the kernel size of the convolution, and the size of the image is , the calculation formulas for the parameters and calculation amount of ordinary convolution and depth convolution are as follows: The parameter formula of ordinary convolution is: (1) The calculation formula of ordinary convolution is: (2) The parameter formula of depth convolution is: (3) The calculation formula of depth convolution is: (4) In order to solve the problem of accuracy degradation caused by DWConv, the SimAM parameter-free attention mechanism is further introduced. SimAM is an attention mechanism that can effectively improve the performance of the model. It adjusts the weights of the feature map in a parameter-free way to enhance the representation ability of key features. The introduction of the SimAM attention mechanism can not only maintain the advantages of parameters and computation brought by DWConv, but also improve the recognition accuracy of the model by accurately capturing important features. Therefore, SimAM can make up for the negative impact of DWConv on accuracy, thereby ensuring that the recognition ability of the model is effectively improved while reducing the computational complexity.

[0035] SimAM (Similarity Attention Module) is a module based on the self-attention mechanism. Its core idea is to strengthen the model's attention to important features by calculating the similarity between samples, thereby improving the model's performance. Compared with the traditional attention mechanism, SimAM does not need to calculate complex weighted matrices or obtain the relationship between features through large-scale matrix operations. It only distributes attention by calculating the similarity between features, thereby reducing the complexity of calculation. This feature makes SimAM more efficient in calculation than some classic self-attention modules.

[0036] In the present invention, SimAM can effectively transfer important information between the layers of the network by focusing on the similarities between rice, and can better mine the potential important details in the image when processing abnormal rice detection. Since the target rice in this study is relatively small, it is relatively difficult to identify, and SimAM can emphasize and process the features more specifically, improving the sensitivity of the model to key areas and local features, thereby improving the recognition effect of rice.

[0037] Many deep learning models usually require a large amount of labeled data to improve their performance during training. SimAM can achieve good performance on smaller data sets by effectively enhancing the similarity relationship between features. This enables it to perform better than other complex models in certain data-scarce situations and has better data efficiency.

[0038] The schematic diagram of the SimAM model is as follows Figure 8As shown in the figure, Generation represents the process of generating attention weights, and Expansion represents the process of expanding or amplifying the generated attention weights. Fusion is a fusion method that fuses feature maps or attention weights from different sources to generate richer feature representations.

[0039] SimAM's contributions to performance improvement mainly include integrating global and local features, capturing the subtle differences and complex structures of input data through 3D attention weighting, making the model more accurate and efficient in processing complex tasks, parameter-free design to simplify the model architecture, reduce computational complexity, improve performance, and be suitable for resource-constrained environments, quantifying the uniqueness of neurons and their relevance through energy functions, and optimizing the attention mechanism. The simplest implementation of finding these neurons is to measure the linear separability between a target neuron and other neurons, based on which the following energy function is defined: (5) in, = and = yes t and The linear transformation of t and is the input feature target neurons and other neurons in a single channel. i is the index in the spatial dimension, M = H×W is the number of neurons on this channel. and is the weight and bias of the transformation. All values ​​in formula (5) are scalars. equal , and all other for When , formula (5) reaches the minimum value, where and are two different values. By simplifying this formula, formula (5) is equivalent to finding the target neuron t and all other neurons in the same channel. For simplicity, binary labels (i.e., 1 and -1) are adopted, and a regularizer is also added to Equation (5). The final energy function is shown as follows: (6) In theory, there are M energy functions for each channel. Solving all these equations through some iterative solver (such as SGD) is computationally heavy. Among them, formula (6) has a and A fast closed-form solution of can be obtained as follows: (7) (8) in, and is the mean and variance calculated on all neurons in this channel except t in this channel. Since the existing solutions shown in Equation (7) and Equation (8) are obtained on a single channel, it is reasonable to assume that all pixels in a single channel follow the same distribution. Given this assumption, the mean and variance can be calculated on all neurons and reused for reporting on all neurons on this channel. It can significantly reduce the computational cost and avoid iteratively calculating µ and σ for each position. Therefore, the minimum energy can be calculated as follows: (9) and Equation (9) shows that the energy The lower the value, the more different the neuron t is from the surrounding neurons and the more important it is to visual processing. Therefore, the importance of each neuron can be calculated by 1 / To obtain. Using the scaling operator to perform feature refinement, the entire optimization phase of this module is: (10) where E is the total number of channels and spatial dimensions. , ⊙ represents element-by-element multiplication. Add the sigmoid function to limit the excessive value in E and compare it with the original feature map Multiply them together to get a weighted feature map. It does not affect the relative importance of each neuron because is a monophonic function.

[0040] Finally, the efficient BiFPN (Bidirectional Feature Pyramid Network) architecture is introduced at the neck of the network. BiFPN is an architecture that optimizes feature fusion through bidirectional information flow, which can significantly improve the fusion effect of multi-scale features. By introducing BiFPN, the model can better transmit information in multi-scale feature maps while maintaining a low amount of computation and parameters, thereby improving the model's sensitivity to rice.

[0041] BiFPN is an efficient network architecture for computer vision tasks, especially for effective feature fusion at different scales in images. The main purpose of BiFPN is to enhance multi-scale feature expression through an efficient feature pyramid architecture. The traditional feature pyramid network (FPN) only performs top-down feature fusion, while BiFPN supports both top-down and bottom-up feature flow by introducing bidirectional connections. This two-way processing can better capture feature information of different scales and improve the utilization efficiency of multi-scale information. BiFPN adopts a weighted feature fusion method. By introducing learnable weights, BiFPN can dynamically adjust the fusion method of features of different scales, thereby avoiding the calculation of overly complex fully connected structures. This method can reduce unnecessary calculations and improve computational efficiency.

[0042] The structure of BiFPN is as follows Fig. 9 As shown in the figure, P3, P4, P5, P6, and P7 represent the output layers of the backbone network. Each output layer has corresponding output features (including the number of channels, feature size, and other information). For example, the output feature size of P3 is the input image resolution / 2^3, the output feature size of P4 is the input image resolution / 2^4, and so on. The output feature size of P7 is the input image resolution / 2^7, which are P3_in, P4_in, ..., P7_in in the figure. The hollow circles without color represent features, the solid circles with color represent operators, and the wired circles represent weights W. The calculation of each operator is shown below: (11) (12) (13) (14) (15) (16) (17) (18) Among them, Resize is usually used for upsampling or downsampling operations for resolution matching. Represents a smaller positive number to prevent the denominator from being zero, ensuring the stability of numerical calculations and the smooth progress of training. Conv is usually a convolution operation used for feature processing. BiFPN can generate more powerful feature representations through multiple cascades and efficient feature fusion. Since BiFPN can more effectively fuse multi-scale features, it can usually improve the accuracy in detection tasks, especially in scenarios where the target scale varies greatly. It can improve the detection performance of small and large objects and improve the overall effect of the model. At the same time, it reduces the need for manual design and adjustment by introducing an adjustable weighting mechanism, and improves the network's automatic learning ability.

[0043] The introduction of BiFPN (Bidirectional Feature Pyramid Network) in Neck of YOLOv11 has the following core advantages compared to the Neck architecture of the original model: Improved feature fusion capability: The neck part of the original model YOLOv11n uses a simple splicing method when fusing features at different levels, which fails to effectively distinguish the importance of features at different scales, which may introduce noise interference or semantic bias. After the introduction of the BiFPN architecture, features at different levels are weightedly fused through learnable weights, and these weights are automatically optimized during the back propagation process. This method can dynamically adjust the contribution of high-level semantic information and low-level detail information, significantly improving the detection accuracy of small targets (such as rice) and occluded targets.

[0044] Multi-scale information flow optimization: The neck part of the original model YOLOv11n only supports one-way information flow, although it supports top-down (high-level → low-level) and bottom-up (low-level → high-level) paths, but this leads to redundant calculations. After the introduction of the BiFPN architecture, the network can achieve two-way cross-scale interaction between high-level and low-level layers, forming a closed-loop feature reuse, enhancing multi-scale expression capabilities, and optimizing computational efficiency through parameter sharing.

[0045] Strong adaptability and support for compound scaling: The neck part of the original YOLOv11n is difficult to flexibly adapt to different hardware scenarios due to its fixed hierarchical structure (for example, edge devices require lower computational workloads). After the introduction of the BiFPN architecture, compound scaling can be achieved, and the depth, width, and input resolution of the neck part can be uniformly adjusted, thereby achieving a balance between model accuracy and computational efficiency to meet the needs of different hardware platforms.

[0046] In summary, the YOLOv11n model is optimized and improved by introducing DWConv, SimAM parameter-free attention mechanism and BiFPN architecture. The combination of these strategies not only effectively reduces the amount of calculation and parameters of the model, but also improves its accuracy and efficiency in rice recognition tasks, ensuring the good performance of the model in practical applications.

[0047] like Figure 6 The figure shows a simplified diagram of the Rice-YOLO-AD model. The information flow in the neck part of the model is divided into four paths: The modules passed through in the first path are convolution module, fusion module and C3K2 module respectively; The second path is divided into two branches after the convolution module; the first branch passes through the Fusion module in the first path and the C3K2 module in the first path in sequence; the second branch passes through the Fusion module, the C3K2 module, the Fusion module in the first path, and the C3K2 module in the first path in sequence; The third path is divided into two branches after the convolution module; the first branch passes through the Fusion module, the C3K2 module, the upsampling module, the Fusion module in the second path, the C3K2 module in the second path, the Fusion module in the first path, and the C3K2 module in the first path; the second branch passes through the Fusion module, the C3K2 module, the convolution module, and then enters the Fusion module in the fourth path; The fourth path is divided into two branches after the convolution module; the first branch passes through the upsampling module and then enters the Fusion module of the first branch in the third path; the second branch passes through the Fusion module and the C3K2 module in turn; The C3K2 module in the first path outputs information into the head network. The C3K2 module in the first branch also inputs information into the Fusion module in the third path through a convolution module. The output information of the C3K2 module in the second branch of the third path enters the head network; The C3K2 module output information of the second branch in the fourth path enters the head network.

[0048] Embodiment 4, This embodiment is a further limitation of Embodiment 1, and step S3 is further described. During the model training process, the input size of the data set is set to 640 × 640, the number of training rounds is 200, the basic learning rate is set to 0.01, the batch size is set to 16, and the optimizer uses SGD. The experiment is deployed on a computer equipped with an Intel(R) Xeon(R) W-2245 CPU (3.9GHz) and an NVIDIA Quadro RTX 5000 GPU (16GB). The operating system is Windows 10, and the software configuration is installed as Anaconda3-2021.11-windows version. The PyCharm compiler is used, and the Pytorch2.1.2 built-in Python3.8.19 programming language is given. All algorithms are run in the same environment.

[0049] Embodiment 5, This example further illustrates the beneficial effects of the Rice-YOLO-AD model proposed in the present invention on abnormal rice detection through specific experimental data.

[0050] 1. First, the model evaluation indicators are explained.

[0051] This embodiment uses indicators such as recall (R), precision (P), F1 score (F1), AP, and mAP to evaluate the performance of the model. Taking the binary classification problem as an example, the actual result is defined as positive and the predicted result is positive, recorded as TP; if the actual result is negative, the predicted result is positive, recorded as FP; if the actual result is positive, the predicted result is negative, recorded as FN; if the actual result is negative, the predicted result is negative, recorded as TN. Precision is the ratio of the number of correctly predicted positive samples to the total number of samples predicted to be positive. Recall is the ratio of the number of correctly identified positive samples to the total number of actual positive samples. The average of the AP values ​​of multiple categories is mAP (average precision), and the higher the value, the higher the average accuracy of the model in detecting each category. The F1 score is the harmonic mean of precision and recall, providing a single indicator that balances the two.

[0052] GFLOPs (Giga Floating-Point Operations Per Second) refers to one billion floating-point operations per second. Weights refers to weights. Postprocess per image refers to the post-processing of each image output by the model.

[0053] 2. Ablation Experiment The specific contents of the ablation experiments in this study are shown in Tables 1 and 2. When deep convolution (DWConv) was introduced alone, the experimental results were consistent with expectations, and the overall recognition effect of the model decreased. However, this change resulted in a 11.34% reduction in the number of model parameters, a 4.76% reduction in GFLOPs (floating point operations per second), a 10.91% reduction in weight, and a 54.72% reduction in the post-processing time per image. This result shows that although DWConv performs well in reducing the amount of computation and model parameters, its negative impact on recognition performance cannot be ignored. When the SimAM module was added alone, the complexity of the model remained unchanged, but its recognition performance was significantly improved. Specifically, mAP50 increased by 4.74%, mAP50-95 increased by 5.42%, recall increased by 5.25%, precision increased by 2.68%, F1 score increased by 3.96%, and the post-processing time per image was reduced by 56.6%. This result further verifies the effectiveness of the SimAM module in improving model performance, indicating that it can significantly enhance recognition accuracy without increasing model complexity. When the DWConv and SimAM modules are introduced at the same time, although the introduction of DWConv is a factor in the decline in recognition accuracy, the introduction of SimAM compensates for this effect to a certain extent. Specifically, mAP50 and mAP50-95 have been improved to varying degrees, while the number of parameters and the amount of calculation of the model remain reduced, indicating that SimAM can effectively balance the accuracy loss caused by DWConv, thereby improving the overall performance. When the BiFPN module is introduced, the experimental results show that mAP50 is improved by 5.25%, mAP50-95 is improved by 5.42%, and the recall rate is improved by 4.1%, while the number of parameters is reduced by 25.55%, the weight is reduced by 23.64%, and the post-processing time of each image is significantly reduced by 60.38%. These results show that the BiFPN module significantly reduces the computational burden and enhances the efficiency of the model while improving recognition accuracy. When these three strategies (DWConv, SimAM and BiFPN) are applied simultaneously, the overall performance of the model is greatly improved. Specifically, mAP50 is improved by 5.27%, mAP50-95 is improved by 7.35%, recall rate is improved by 4.72%, and F1 score is improved by 1.41%, while the number of model parameters is reduced by 38.86%, GFLOPs is reduced by 4.76%, weight is reduced by 34.55%, and post-processing time of each picture is reduced by 54.72%. This combined strategy not only significantly improves the recognition accuracy of the model, but also greatly reduces the computational cost while maintaining lightweight and high efficiency. In summary, by reasonably introducing these modules, the method of the present invention effectively improves the recognition accuracy of the model for rice, while optimizing its computational performance, achieving a balance between high efficiency and high precision.

[0054] Table 1:

[0055] Table 2:

[0056] The convergence curve of the Rice-YOLO-AD model during training is as follows Fig.10 As shown in the figure, train / box_loss: the bounding box regression loss on the training set.

[0057] train / cls_loss: classification loss on the training set.

[0058] train / dfl_loss: distribution loss on the training set.

[0059] val / box_loss: bounding box regression loss on the validation set.

[0060] val / cls_loss: classification loss on the validation set.

[0061] val / dfl_loss: distribution loss on the validation set.

[0062] metrics / precision (B): Precision, which measures the proportion of samples predicted to be positive that are actually positive.

[0063] Metrics / recall (B): Recall rate, which measures the proportion of samples that are correctly predicted to be positive among all samples that are actually positive.

[0064] metrics / mAP50 (B): Average precision (AP) at an IoU threshold of 0.5.

[0065] metrics / mAP50-95 (B): The average mean precision over IoU thresholds from 0.5 to 0.95.

[0066] It can be observed from the figure that the recognition performance of the model is poor in the early stage of training, and it is difficult to accurately identify the target. The emergence of this phenomenon is mainly related to the characteristics and size of the target object. At the beginning of training, the model faces a relatively small target with relatively simple features, rice, so it is difficult for the model to effectively identify the target in the early stage. Specifically, due to the small size of rice and the small number of samples, the model has large errors in the feature extraction and target positioning process, which affects the accuracy of recognition. However, with the gradual increase in the number of training rounds, the recognition effect of the model has shown a significant improvement. This improvement is mainly due to the continuous optimization of the model during the training process and the gradual adjustment of weights, so that the network can better adapt to the complex feature space. Through back propagation, the model gradually reduces the prediction error and optimizes its feature extraction ability. Especially when the target object is small, the model can better focus on the target of rice and gradually improve its recognition accuracy. From the changes in the loss curve and the accuracy curve, the model gradually tends to be stable during the training process. The loss curve drops rapidly in the early stage of training, indicating that the model error is constantly decreasing and the prediction results are gradually approaching the true label. At the same time, the mAP50 curve also rose rapidly, reflecting that the prediction accuracy of the model gradually improved and eventually stabilized. This phenomenon shows that after multiple rounds of iterations, the model gradually completed the data fitting process and finally achieved a relatively ideal recognition effect on the training set and validation set. In addition, as the number of training rounds increases, the model is gradually able to recognize more detailed features and show stronger generalization ability when dealing with small targets. In general, as the training deepens, the model is continuously optimized and adjusted, and finally completes the relatively accurate recognition of rice after convergence.

[0067] 3. Comparative Experiment 1. Comparison with classic target detection models In order to verify the performance advantage of Rice-YOLO-AD in the abnormal rice detection task, this embodiment is compared with multiple classic target detection models (including YOLOv6n, YOLOv8n, YOLOv10n, YOLOv10s, YOLOv10m, YOLOv11n and YOLOv12n). The experimental results are shown in Table 3. The experimental results show that Rice-YOLO-AD has achieved 95.09% in the mAP50 index, which is an excellent performance. Specifically, although the mAP50-95 of YOLOv10m is 0.63% higher than that of Rice-YOLO-AD, Rice-YOLO-AD performs better than YOLOv10m in other key recognition indicators. More importantly, the model parameters of YOLOv10m are 9.39 times that of Rice-YOLO-AD, the GFLOPs are 9.82 times, and the weights are 9.31 times that of Rice-YOLO-AD, showing its significant consumption of computing resources. Therefore, although YOLOv10m performs slightly better in some indicators, its high computational overhead significantly reduces its efficiency in practical applications. In addition, YOLOv8n also performs well in recognition performance, and its mAP50 is 1.14% higher than the original model YOLOv11n. However, compared with Rice-YOLO-AD, although YOLOv8n has a slight advantage in accuracy, it does not show obvious advantages in other aspects. In particular, Rice-YOLO-AD shows a more superior cost-effectiveness in the control of computing resources and parameter quantity. Therefore, considering all indicators comprehensively, Rice-YOLO-AD not only performs well in recognition accuracy, but also has significant advantages in the model's computational efficiency and optimization of parameter quantity. In summary, Rice-YOLO-AD achieves a good balance between performance and efficiency, especially in the rice target detection task. Compared with other classic target detection models, it can significantly reduce the consumption of computing and storage resources while ensuring high recognition accuracy, showing excellent application potential.

[0068] Table 3:

[0069] like Fig.11As shown in the figure, in the comparative experiment with the classic target detection model, the performance of the network model Rice-YOLO-AD proposed in this application on the mAP50 curve shows the best convergence effect. Specifically, the mAP50 curve of Rice-YOLO-AD shows a rapid and steady upward trend during the training process, indicating that the model can significantly improve the accuracy of target detection in a shorter training time and can quickly tend to the best performance. This phenomenon shows that the model effectively optimizes the feature extraction and decision-making mechanism during the learning process, especially in the early stages of training. Through the carefully designed network architecture and optimization strategy, the detection accuracy of the model has been rapidly improved, and then convergence has been achieved. Compared with other classic comparison models, the performance of Rice-YOLO-AD on the mAP50 curve is particularly stable and has small fluctuations, which shows that the model has shown strong recognition ability and superior generalization performance in rice detection tasks. In contrast, YOLOv12n performed relatively poorly in the same experiment, especially in the early stages of training, where its mAP50 curve showed a significant drop in the rapid rise stage, indicating that it faced great challenges in the optimization process, resulting in the failure to continuously improve the recognition effect of the model during training. This further confirms the effectiveness of Rice-YOLO-AD's optimization strategy and model design in solving rice detection tasks.

[0070] 2. Comparative experiments with different backbone networks In the field of object detection, the backbone network is usually the core part of feature extraction, and its role is to extract meaningful feature information from the input image. These feature information will then be passed to the subsequent network layers for target positioning and classification. Therefore, for the target detection task, the choice of the backbone network directly affects the detection accuracy, computational efficiency and model complexity. In order to verify the advantages of the network model Rice-YOLO-AD, seven classic classification backbone networks (including RepHGNetV2, EfficientVit, FasterNet_T0, MobileNetV4_s, UniReplkNet_S, SwinTransformer and DynamicHGNetV2) were systematically compared based on the YOLOv11n model. The comparison results are shown in Table 4. In these comparative experiments, the SwinTransformer network showed the best recognition effect. Specifically, compared with Rice-YOLO-AD, SwinTransformer is 1.61% higher in mAP50, 3.94% higher in accuracy, and 1.17% higher in F1 score. However, although SwinTransformer has achieved significant performance improvements, its cost cannot be ignored. The number of parameters of SwinTransformer is 18.22 times that of Rice-YOLO-AD, the model weight is 16.67 times that of Rice-YOLO-AD, and the computation time is 3.52 times that of Rice-YOLO-AD. This result fully reflects the trade-off between model performance and computational cost in the selection of the backbone network. Through these comparative experiments, the advantages of Rice-YOLO-AD in computational efficiency and number of parameters are verified.

[0071] Table 4:

[0072] 3. Comparison with other attention mechanisms In order to verify the superiority of the attention mechanism SimAM used in the present invention, this experiment compared seven classic attention modules, including CPCA, AFGCAttention, CAFM, MPCA, DAttention, TripletAttention and MLCA, and the verification results are shown in Table 5. By comparing the performance of these models in terms of average recognition accuracy, CAFM performed the best among all modules, and its mAP50 value was 0.56% higher than SimAM. However, this advantage is not without cost. Specifically, the number of parameters of the CAFM model increased by 13.4% compared with the original model, GFLOPs increased by 4.76%, and the weight also increased by 12.73% accordingly. This shows that although CAFM has improved in accuracy, it has paid a great price in terms of computational complexity and model volume. In contrast, SimAM maintains a low computational cost and model complexity while improving accuracy. Specifically, SimAM significantly reduces the image post-processing time by 56.6% without adding additional parameters, GFLOPs and weights. In addition, the mAP50 value of SimAM is 3.89% higher than that of the original model. This result shows that SimAM can effectively improve the recognition accuracy of the model while maintaining low computing resource consumption, especially without significantly increasing the computational complexity. By comparing the experimental results, it can be clearly seen that SimAM can outperform other classic attention mechanism modules in many aspects without significantly increasing the computational burden, demonstrating its wide applicability and potential in the abnormal rice detection task.

[0073] Table 5:

[0074] 4. Comparison of different positions of attention mechanism In the object detection algorithm, the attention mechanism improves the performance of the model by focusing on important areas or features in the image. Integrating the attention mechanism into different positions in the backbone network of the object detection algorithm will have different effects on the performance of the model. This is because the attention mechanisms at different positions act on different feature layers, affecting the extraction and fusion of features, and ultimately affecting the effect of object detection.

[0075] Therefore, in order to verify the rationality of the placement of the attention mechanism in the present invention, the following comparative experiments were conducted: Fig.12As shown in Figure 6, SimAM is placed in five different positions in the YOLOv11 backbone to test the recognition effect. The specific results are shown in Table 6. It is worth noting that the further the SimAM attention mechanism is placed, the better its mAP50. This is mainly because the features of the previous network layer are relatively coarse and contain a lot of detailed information. As the number of network layers increases, the abstraction level of the features increases and the information becomes more compact. At this time, placing SimAM can better focus on meaningful global information, avoid excessive feature selection or weighting in the early stage, and reduce the loss of useful information. Since the feature maps of the back-end network are usually smaller than those of the front-end and the amount of calculation is lower, placing the attention mechanism in these layers will not introduce too much computational burden, but can produce a significant improvement in accuracy. At the same time, adding an attention mechanism to the front of the network backbone may lead to unnecessary weighting of low-level features, which may have little effect on the final target recognition in target detection, but may increase the computational overhead. In contrast, placing SimAM at the back of the network can avoid this redundant calculation and focus more on features that have a greater impact on the final target positioning and classification. This result also further verifies the superiority of the present invention's implementation scheme for the attention mechanism.

[0076] Table 6:

[0077] 4. The impact of image size on the model In the target detection algorithm, the choice of image size has an important impact on the performance. Different image sizes will affect the algorithm's computational efficiency, accuracy, and model training effect. Therefore, when designing and applying the target detection algorithm, it is necessary to weigh the impact of image size. Therefore, in order to verify the superiority of the image size of 640×640 used in the present invention, this study compared the effects of 7 different image sizes on the model recognition results. The specific contents are shown in Table 7, among which the image size of 640×640 obtained the best recognition accuracy, which further reflects the superiority of the design of the present invention.

[0078] Since rice is a relatively small object in the image, the smaller the image size, the less the model can recognize small objects or objects with more details, resulting in a weaker rice recognition effect. Because in the Rice-YOLO-AD network, a smaller image size will lead to a lower resolution of the feature map, which will affect the network's ability to learn rice details. In addition, the network's convolution operation will have a greater impact on the loss of low-resolution image details, reducing accuracy. Larger input images will retain more detailed information, especially for those smaller objects, increasing the image size will help improve detection accuracy.

[0079] Table 7:

[0080] 5. Demonstration of identification effects of different blending methods like Figure 13-15 As shown in the figure, the recognition effect before and after the model improvement was tested. The top picture shows the detection result of YOLOv11n, and the bottom picture shows the detection result of Rice-YOLO-AD model. Normal means complete and edible rice, and the subsequent values ​​indicate the confidence level. Broken means broken rice, mould means moldy rice, and dirty means contaminated and discolored rice. Fig.13 and Fig.14 The improved model significantly improved the confidence of the overall recognition effect, especially in the recognition of broken rice, where the accuracy was significantly improved. Fig.15 In the example above, the original model failed to correctly distinguish broken and moldy rice, and mistakenly identified it as normal rice. However, the Rice-YOLO-AD model was able to accurately identify broken and moldy rice, showing higher recognition ability. Overall, the Rice-YOLO-AD model showed stronger recognition ability when dealing with different types of rice, especially in mixed and dense situations, the recognition accuracy was significantly improved. This shows that the Rice-YOLO-AD model plays a key role in enhancing detection effects and improving accuracy.

Claims

1. A lightweight target detection method for abnormal rice detection, characterized in that: The method comprises the following steps: S1. Dataset construction: Collect different types of abnormal rice images, label them, and divide them into training sets, validation sets, and test sets; S2. Model construction: Based on the YOLOv11n model, the YOLOv11n model is improved in combination with the morphological characteristics of rice to build a model suitable for abnormal rice detection; The improvements to the YOLOv11n model are as follows: Reduce the number of model parameters and computational complexity; Introducing the attention mechanism; Optimize information fusion architecture; S3, model training: using the constructed data set to train the model suitable for abnormal rice detection, and adjusting the model parameters until the model meets the detection requirements; S4. Use the trained model suitable for abnormal rice detection to perform abnormal rice detection.

2. The lightweight target detection method for abnormal rice detection according to claim 1, characterized in that: The different types of abnormal rice include four types, namely broken rice, contaminated and discolored rice, intact edible rice, and moldy rice.

3. The lightweight target detection method for abnormal rice detection according to claim 2, characterized in that: The method of reducing the number of parameters and the amount of calculation of the model is specifically as follows: the fifth convolution module starting from the input in the YOLOv11n model backbone network is replaced with a deep convolution module.

4. The lightweight target detection method for abnormal rice detection according to claim 3, characterized in that: The introduction of the attention mechanism is specifically as follows: adding a SimAM module after the C2PSA module in the YOLOv11n model backbone network.

5. The lightweight target detection method for abnormal rice detection according to claim 4, characterized in that: The optimized information fusion architecture specifically includes: introducing the BiFPN architecture into the neck part of the YOLOv11n model, and changing the information fusion method of the neck part of the YOLOv11n model.

6. The lightweight target detection method for abnormal rice detection according to claim 5, characterized in that: The information fusion method of the neck part of the YOLOv11n model is changed specifically as follows: S61, connect the first C3K2 module of the YOLOv11n model trunk from the input to the neck part; S62. Add a convolution module between the modules connecting the trunk and the neck of the YOLOv11n model. S63, replace the Concat module in the neck part with the Fusion module, and perform weighted fusion of features at different levels through learnable weights; S64. Change the one-way information flow mode of the neck part of the YOLOv11n model into a two-way cross-scale interaction mode.

7. The lightweight target detection method for abnormal rice detection according to claim 6, characterized in that: The information flow of the neck part of the model suitable for abnormal rice detection is divided into four paths: The modules passed through in the first path are convolution module, fusion module and C3K2 module respectively; The second path is divided into two branches after the convolution module; the first branch passes through the Fusion module in the first path and the C3K2 module in the first path in sequence; the second branch passes through the Fusion module, the C3K2 module, the Fusion module in the first path, and the C3K2 module in the first path in sequence; The third path is divided into two branches after the convolution module; the first branch passes through the Fusion module, the C3K2 module, the upsampling module, the Fusion module in the second path, the C3K2 module in the second path, the Fusion module in the first path, and the C3K2 module in the first path; the second branch passes through the Fusion module, the C3K2 module, the convolution module, and then enters the Fusion module in the fourth path; The fourth path is divided into two branches after the convolution module; the first branch passes through the upsampling module and then enters the Fusion module of the first branch in the third path; the second branch passes through the Fusion module and the C3K2 module in turn; The C3K2 module in the first path outputs information into the head network. The C3K2 module in the first branch also inputs information into the Fusion module in the third path through a convolution module. The output information of the C3K2 module in the second branch of the third path enters the head network; The C3K2 module output information of the second branch in the fourth path enters the head network.

8. A lightweight target detection system for abnormal rice detection, characterized in that: The system includes the following modules: Module for constructing data sets: collect different types of abnormal rice images, label them, and divide them into training sets, validation sets, and test sets; Model building module: Based on the YOLOv11n model, the YOLOv11n model is improved in combination with the morphological characteristics of rice to build a model suitable for abnormal rice detection; The improvements to the YOLOv11n model are as follows: Reduce the number of model parameters and computational complexity; Introducing the attention mechanism; Optimize information fusion architecture; Model training module: Use the constructed data set to train the model suitable for abnormal rice detection and adjust the model parameters until the model meets the detection requirements; A module for abnormal rice detection using a trained model suitable for abnormal rice detection.

9. A computer device comprising: A processor and a memory, characterized in that the memory is used to store executable instructions of the processor, and the processor is configured to execute the lightweight target detection method for abnormal rice detection described in any one of claims 1 to 7 by executing the executable instructions.

10. A computer storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is run, the lightweight target detection method for abnormal rice detection according to any one of claims 1 to 7 is executed.

Images