Lightweight semi-supervised target detection method based on joint estimation and scale fusion

Abstract

The invention belongs to the technical field of artificial intelligence, and particularly relates to a lightweight semi-supervised target detection method based on joint estimation and scale fusion, which introduces dense learning knowledge, a joint confidence estimation module and a mixed scale feature pyramid module, and replaces a traditional sparse pseudo-frame with pixel-level prediction. A continuous value output by a network is directly used as a pseudo label, so that a complicated post-processing step is avoided; the joint confidence estimation module adopts a double-branch structure, and performs confidence estimation on classification and positioning quality respectively, so that the accuracy of pseudo labels is improved; and the mixed scale feature pyramid module adaptively fuses different scale features, so that the robustness of the model to scale change is enhanced. Through combined action of a series of innovative designs, the performance of semi-supervised target detection is remarkably improved, and powerful support is provided for practical application of the technology.

Description

Technical Field

[0001] The present invention belongs to the technical field of artificial intelligence, and in particular relates to a lightweight semi-supervised target detection method based on joint estimation and scale fusion. Background Art

[0002] With the development of deep learning technology, object detection has made significant progress in the field of computer vision. However, current object detection methods usually rely on a large amount of labeled data to train the model, which not only requires a lot of manpower, material resources and time costs, but also in some practical application scenarios, it may be difficult or even impossible to obtain enough labeled data. Therefore, semi-supervised object detection has emerged as an effective solution, which improves model performance by utilizing a small amount of labeled data and a large amount of unlabeled data, thereby reducing the need for expensive labeled data.

[0003] Existing one-stage anchor-free detectors face several key problems when dealing with semi-supervised object detection tasks. First, the quality of pseudo-labels is often not high. This is because traditional methods use sparse pseudo-boxes as pseudo-labels, and their generation process involves complex post-processing steps (such as non-maximum suppression (NMS) and thresholding). These steps increase the computational complexity and introduce many hyperparameters, whose settings directly affect the quality of pseudo-labels. Secondly, when generating pseudo-labels, the mismatch between the classification score and the positioning quality of the one-stage detector leads to inaccurate pseudo-labels generated by the teacher model, which in turn affects the learning effect of the student model. Finally, for targets with large scale variations, existing methods are insufficient in their ability to adapt to different scales, especially for the detection of small targets, which are prone to false detection or missed detection.

[0004] Problems faced by one-stage anchor-free detectors in semi-supervised object detection include low pseudo-label quality, redundant post-processing steps, and poor scale adaptability, which limit the overall performance and practicality of the model. Summary of the invention

[0005] The purpose of the present invention is to provide a lightweight semi-supervised target detection method based on joint estimation and scale fusion to solve the problems raised in the above background technology.

[0006] To achieve the above object, the present invention adopts the following technical solution: a lightweight semi-supervised target detection method based on joint estimation and scale fusion, comprising the following steps: Using labeled image data and unlabeled image data as input, a teacher model is used to generate dense pseudo labels for unlabeled images based on pixel-level predictions; Using a joint confidence estimation module to classify and estimate the positioning quality of the generated dense pseudo-labels, generate a joint confidence score, perform region division according to the joint confidence score, and divide the input image into a learning region and a suppression region; Construct a mixed-scale feature pyramid, including a large-scale feature pyramid in a conventional view and a small-scale feature pyramid in a downsampled view, and fuse the features of each scale through a feature fusion module to generate a mixed-scale feature pyramid; The generated mixed-scale feature pyramid is used for training the student model to optimize the model parameters, and finally an optimized semi-supervised target detection model is obtained.

[0007] Preferably, generating dense pseudo labels for unlabeled images based on pixel-level predictions by a teacher model comprises: Make pixel-level predictions for unlabeled images and assign target class probabilities to each pixel; Use the continuous values ​​output by the network as pseudo labels.

[0008] Preferably, the joint confidence estimation module is configured as a dual-branch structure, comprising: The classification branch is responsible for identifying the object category and outputting the classification score; The localization branch is used to estimate the localization quality and output the predicted intersection-over-union ratio; The outputs of the classification branch and the localization branch are combined through a sigmoid function to form a joint confidence score of the detection result.

[0009] Preferably, the performing of region division according to the joint confidence score includes: The input image is divided into a learning area and a suppression area. The learning area contains the first k% pixels with the highest scores, and the rest of the area is set to 0. The learned regions are used to generate high-quality pseudo labels, and the suppressed regions are used to reduce the interference of low-scoring predictions in background regions.

[0010] Preferably, the constructing of a mixed-scale feature pyramid includes: Constructing large-scale feature pyramids in regular views; Constructing small-scale feature pyramids in the downsampled views.

[0011] Preferably, the feature fusion module includes squeeze and excitation operations; The squeeze operation is used to perform global average pooling on the input feature map to generate channel descriptors; The excitation operation is used to learn the correlation between channels through two fully connected layers and generate weight values ​​for each channel. The weight values ​​are normalized by the Sigmoid activation function.

[0012] Preferably, the training of the student model includes: Use labeled data and generated pseudo labels as supervision information; Model parameter optimization is performed through the weighted sum of supervised loss and unsupervised loss; The training process is repeated until the model converges.

[0013] Preferably, the criteria for judging the convergence of the model include but are not limited to: The model's performance on the validation set reaches a preset threshold; The loss function of the model no longer decreases significantly within a certain number of iterations.

[0014] Preferably, the teacher model and the student model both adopt a one-stage anchor-free detector architecture, including: The teacher model is used to generate high-quality dense pseudo labels; The student model is used to learn the pseudo labels generated by the teacher model.

[0015] Preferably, the classification branch adopts the FocalLoss focal loss function, and the positioning branch adopts the binary cross entropy loss function.

[0016] Technical effects and advantages of the present invention: Compared with the prior art, the lightweight semi-supervised target detection method based on joint estimation and scale fusion proposed in the present invention has the following advantages: The present invention introduces dense learning knowledge, joint confidence estimation module and mixed scale feature pyramid module, replaces the traditional sparse pseudo-frame with pixel-level prediction, directly uses the continuous value output by the network as the pseudo-label, and avoids complicated post-processing steps; the joint confidence estimation module adopts a dual-branch structure to estimate the confidence of classification and positioning quality respectively, which improves the accuracy of pseudo-label; the mixed scale feature pyramid module adaptively fuses different scale features, and enhances the robustness of the model to scale changes. This series of innovative designs work together to significantly improve the performance of semi-supervised object detection and provide strong support for the practical application of this technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 A flowchart of a lightweight semi-supervised target detection method based on joint estimation and scale fusion according to the present invention; Figure 2 It is the principle topological diagram of the present invention; Figure 3 It is a flow chart of the training method of the present invention; Figure 4 This is a flow chart of semi-supervised training of the present invention; DETAILED DESCRIPTION

[0018] The following will be combined with the accompanying drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. The specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0019] Before introducing a lightweight semi-supervised target detection method based on joint estimation and scale fusion in this embodiment, it is necessary to explain the meanings of the terms in this embodiment, as follows: Object detection: Object detection is a task in the field of computer vision that aims to identify the location (usually represented by a bounding box) and category of a specific target in an image or video. It analyzes the input data and uses deep learning and other technologies to extract features to determine the existence, location, and category information of the target in the image. It is widely used in security monitoring, autonomous driving, image editing and other fields.

[0020] Two-stage object detector: A two-stage object detector first generates candidate regions that may contain objects through methods such as region proposal networks (RPNs), and then performs more accurate object classification and localization on these candidate regions. FasterRCNN is a common two-stage object detector. Its first stage generates candidate regions, and the second stage classifies and regresses these regions. Compared with a one-stage detector, it may have an advantage in accuracy, but the computational complexity is usually higher and the detection speed is relatively slow.

[0021] One-stage object detector: A one-stage object detector is a detection model that directly predicts objects on an image. It does not need to generate candidate regions like a two-stage object detector, but directly predicts the object category and position at each position in the image. For example, FCOS is a typical one-stage object detector, which is characterized by relatively fast calculation speed, but may have a certain gap with the two-stage detector in terms of positioning accuracy.

[0022] Anchored object detector: Anchored object detector predefines a series of anchor boxes of fixed size and ratio during the detection process. These anchor boxes are distributed on the image as prior knowledge, and the model performs classification and regression adjustment based on the matching degree between the anchor box and the target. For example, FasterRCNN is an anchored object detector. The setting of anchor boxes helps the model to quickly locate the target, but the design and selection of anchor boxes requires certain experience and tuning, and may be limited by the size and ratio of the anchor boxes.

[0023] Anchor-free object detector: Anchor-free object detectors do not rely on pre-defined anchor boxes, but directly classify and locate objects based on feature points or feature regions. For example, FCOS is an anchor-free object detector that determines the location of an object by predicting information such as the center of the object, avoiding some problems caused by anchor boxes, such as the complexity of anchor box matching and the inflexibility of anchor box size, but may face challenges when dealing with some complex scenes.

[0024] Semi-supervised learning: Semi-supervised learning is a learning paradigm in machine learning that combines the characteristics of supervised learning and unsupervised learning. In semi-supervised learning, the training data contains both a small amount of labeled data (with labeled information) and a large amount of unlabeled data (without labeled information). The model improves its performance by using the knowledge learned from labeled data and the information mined from unlabeled data (such as data distribution, structure, etc.), so that it can still achieve good learning results under limited labeled data conditions. It is suitable for scenarios where the cost of obtaining labeled data is high or it is difficult to obtain a large amount of labeled data.

[0025] Semi-supervised object detection: Semi-supervised object detection is the application of semi-supervised learning in the field of object detection. It aims to use a small number of labeled images and a large number of unlabeled images to train the object detection model to improve the performance of the model in the object detection task. By making reasonable use of unlabeled data, such as generating pseudo labels, the model is assisted in learning the characteristics and patterns of the target, solving the problem of limited model performance caused by insufficient labeled data. In practical applications, it can save labeling costs and improve the efficiency and accuracy of object detection.

[0026] Pseudo-label: Pseudo-label is a temporary label assigned to unlabeled data through model prediction in semi-supervised learning. In semi-supervised object detection, a pre-trained model is usually used to predict unlabeled images, and the parts of the prediction results with higher confidence are used as pseudo-labels, which are then used together with the labeled data to train the model. For example, the teacher model predicts the target category and position in the unlabeled image, and the predictions with high confidence are selected as pseudo-labels for the student model to learn. However, the accuracy of pseudo-labels may be affected by many factors, such as the performance of the model itself, data distribution, etc.

[0027] Consistency regularization: Consistency regularization is a technical method in semi-supervised learning. In semi-supervised object detection, for the same unlabeled image, under different forms of data augmentation (such as random cropping, flipping, etc.), the model regularizes the consistency of its output. That is, the model's prediction results for the same unlabeled image after different augmentations should be consistent. In this way, the information in the unlabeled data is used to improve the generalization ability and stability of the model, reduce the model's dependence on labeled data, and enable the model to learn the intrinsic characteristics and patterns of the data.

[0028] This paper focuses on the field of semi-supervised target detection. Aiming at the problems faced by one-stage anchor-free detectors in semi-supervised target detection, such as low pseudo-label quality, redundant post-processing steps, and large impact of scale changes, an innovative semi-supervised target detection method is proposed. By introducing dense learning knowledge, joint confidence estimation module, and mixed-scale feature pyramid module, the pseudo-label quality is effectively improved, the model training process is simplified, and the model's robustness to scale changes is enhanced, thereby improving the performance of semi-supervised target detection. The details are as follows: The present invention provides Figure 1-Figure 4 A lightweight semi-supervised target detection method based on joint estimation and scale fusion is shown, comprising the following steps: Step 1: Using labeled image data and unlabeled image data as input, the teacher model generates dense pseudo labels for unlabeled images based on pixel-level prediction, including: performing pixel-level prediction on unlabeled images and assigning target category probability to each pixel; using the continuous value output by the network as the pseudo label; Labeled data provides clear supervision information for the model to learn the accurate features and categories of the target. Unlabeled data contains rich potential information, and its auxiliary value for model training can be mined through the technical means of the present invention. For example, in actual application scenarios, labeled data may be a small number of images that have been carefully annotated by humans, containing the category and accurate location information of the target; unlabeled data is a large number of unlabeled images of the same type or related scenes.

[0029] The teacher model generates dense pseudo labels for unlabeled images based on pixel-level predictions. Different from traditional pseudo-box-based methods, the present invention directly uses the continuous values ​​output by the network as the pseudo label form.

[0030] The teacher model analyzes the unlabeled image and assigns corresponding prediction information to each pixel, such as the possibility of the target category. The advantage of this is that it avoids the complex post-processing steps in the traditional pseudo-frame method, such as non-maximum suppression (NMS), thresholding, and label assignment. In the traditional method, NMS is used to remove overlapping detection frames, but a threshold needs to be set, and different thresholds will have a significant impact on the results; thresholding and label assignment also require fine-tuning of parameters, and the process is cumbersome. The dense pseudo-label generation method of the present invention cancels these post-processing steps, eliminates the influence of post-processing hyperparameters on the model, and retains richer information, providing a more comprehensive and accurate foundation for the subsequent learning of the student model.

[0031] The teacher model fully utilizes the feature extraction capability of the deep learning network when generating pixel-level dense pseudo-labels on unlabeled images. The network analyzes each pixel in the image and predicts the probability or other relevant information of each pixel belonging to different target categories based on the local features around the pixel and its association with the global features.

[0032] For example, for a complex image containing multiple targets, the teacher model can determine the target category that each pixel may belong to, such as pedestrians, vehicles, buildings, etc. Compared with the traditional pseudo-frame method, this pixel-level prediction method does not require pre-framing and post-processing of the target, avoiding information loss and error introduction caused by inaccurate framing or improper post-processing. At the same time, because the continuous values ​​output by the network are directly used as pseudo-labels, more detailed information is retained, allowing the student model to obtain richer learning resources, which helps to more accurately understand the characteristics and distribution of the target.

[0033] Step 2: using a joint confidence estimation module to classify the generated dense pseudo-labels and estimate the positioning quality, generating a joint confidence score, performing region division according to the joint confidence score, and dividing the input image into a learning region and a suppression region; Furthermore, the joint confidence estimation module is set as a dual-branch structure, including: a classification branch, which is responsible for identifying the object category and outputting the classification score; a localization branch, which is used to estimate the localization quality and output the predicted intersection-over-union ratio; the outputs of the classification branch and the localization branch are combined through a sigmoid function to form a joint confidence score of the detection result. In addition, the classification branch adopts the FocalLoss focal loss function, and the localization branch adopts the binary cross entropy loss function.

[0034] The above-mentioned region division according to the joint confidence score includes: dividing the input image into a learning area and a suppression area, the learning area contains the top k% pixels with the highest scores, and the other areas are set to 0; the learning area is used to generate high-quality pseudo labels, and the suppression area is used to reduce the interference of low-score predictions in the background area.

[0035] Specifically, the joint confidence estimation module adopts a dual-branch structure, where one branch is responsible for identifying the object category and outputting the classification score, and the other auxiliary branch is used to estimate the localization quality and output the predicted IoU (Intersection over Union).

[0036] For example, when detecting a vehicle, the classification branch determines whether the object in the image is a vehicle and outputs the corresponding classification score; the localization branch calculates the IoU with the real annotation based on the position and shape of the vehicle in the image. The outputs of the two branches are combined through the sigmoid function to form a joint confidence score for the detection result.

[0037] In the dual-branch structure of the joint confidence estimation module, the classification branch adopts a suitable classification loss function, such as the FocalLoss focal loss function, to optimize the learning of the classification task.

[0038] FocalLoss can effectively alleviate the problem of class imbalance, making the model pay more attention to difficult-to-classify samples and improving classification accuracy. The IoU loss of the auxiliary branch uses a two-process cross entropy loss function to accurately measure the difference between the predicted IoU and the true IoU.

[0039] During the training process, the two branches work together to jointly optimize the model parameters. For example, in the initial stage, the classification branch may learn the approximate category of the target quickly, but the localization branch is relatively slow. As the training progresses, the two branches influence each other, and the improvement of the localization branch helps the classification branch further optimize the judgment of the target boundary, thereby improving the joint confidence.

[0040]

[0041] FocalLoss focal loss function, where Indicates the probability that the sample predicted by the model belongs to the true category, and are two adjustable hyperparameters.

[0042]

[0043] Binary cross entropy loss function, where Indicates the probability that the sample predicted by the model is a positive sample.

[0044] Based on this joint confidence score, the input image is divided into a learning region and a suppression region. The region division strategy is dynamically adjusted according to the joint confidence score. For different images and target scenes, the learning region can be adaptively determined to ensure that the model focuses on the most valuable information for learning.

[0045]

[0046]

[0047] Where C is the number of categories, is the score of the corresponding category, represents the maximum score predicted by the teacher across all categories.

[0048] The learning area contains the top k% pixels with the highest scores, which are considered to be the areas most likely to contain targets or make important contributions to target detection, while other areas are suppressed to 0. This region division strategy can effectively screen out high-quality learning areas and reduce the interference of low-score predictions in background areas, thereby improving the accuracy of pseudo labels and enabling the student model to focus on key information during the learning process, thereby improving learning efficiency.

[0049] Step 3: Construct a mixed-scale feature pyramid, including a large-scale feature pyramid in a conventional view and a small-scale feature pyramid in a downsampled view, and fuse the features of each scale through a feature fusion module to generate a mixed-scale feature pyramid; Specifically, constructing a mixed-scale feature pyramid includes: constructing a large-scale feature pyramid in a regular view; and constructing a small-scale feature pyramid in a downsampled view.

[0050] This module constructs a large-scale feature pyramid in the conventional view and a small-scale feature pyramid in the downsampled view. During the construction process, a 0.5 times downsampling rate is used so that The characteristics and Sharing the same spatial size facilitates subsequent feature fusion.

[0051] For example, when processing high-resolution images, the large-scale feature pyramid can capture the overall features and detail information of the target, while the small-scale feature pyramid focuses on the feature changes of the target at different scales. The feature fusion module adaptively fuses features of each scale to generate a mixed-scale feature pyramid. The fusion module first performs a squeeze operation, that is, it performs global average pooling on the input feature map, generates a channel descriptor, compresses the spatial information of the feature map, and highlights the feature differences between channels. Then an excitation operation is performed, using two fully connected layers to learn the correlation between channels and generate weight values ​​for each channel. The weight values ​​are normalized by the Sigmoid activation function as the weight of the linear combination of features. In this way, the model can adaptively fuse features of different scales according to their importance, enhance the robustness to scale changes, and better adapt to the detection needs of targets of different sizes;

[0052] X i is the original image, P i are feature maps of different levels, γ is used as the weight of the linear combination of features, and these features are adaptively fused through the feature fusion module to finally generate a mixed-scale feature pyramid.

[0053] Furthermore, the feature fusion module includes squeeze and excitation operations; the squeeze operation is used to perform global average pooling on the input feature map to generate a channel descriptor; the excitation operation is used to learn the correlation between channels through two fully connected layers and generate a weight value for each channel, and the weight value is normalized by a Sigmoid activation function.

[0054] The fusion module first performs a squeeze operation, globally average pools and compresses the feature map, and then performs an excitation operation. The compressed vector passes through two fully connected layers to learn the correlation between channels (feature maps) and generates a weight value for each channel. The weight value is normalized by the Sigmoid activation function as the weight of the linear combination of features.

[0055] Squeeze: Perform global average pooling on the input feature map to generate channel descriptors.

[0056] Motivation: A simple self-gating mechanism is used to capture the dependencies between channels.

[0057]

[0058]

[0059]

[0060] Among them, H, W, C are the height, width, and number of category channels of the feature map, δ is the ReLU function, W1 and W2 are the weight parameters of the fully connected layer, and σ is the sigmoid activation function.

[0061] When constructing a mixed-scale feature pyramid, the features at each level of the large-scale feature pyramid and the small-scale feature pyramid have different semantic information and spatial resolution. Large-scale features contain more detailed information and are suitable for accurately locating targets; small-scale features have stronger semantic information and are helpful for classifying targets.

[0062] For example, when detecting small targets (such as traffic signs in the distance), small-scale feature pyramids can provide more abstract and representative features to help the model quickly identify the target category; When accurately framing the target, the detailed information of the large-scale feature pyramid plays a key role. The squeezing operation in the feature fusion module compresses the feature map of each channel into a single value through global average pooling to obtain a channel descriptor, which reflects the overall characteristics of the channel.

[0063] The excitation operation establishes a nonlinear relationship between channels through two fully connected layers and the Sigmoid activation function, assigning different weights to the features of different channels, so that the model can perform adaptive fusion according to the importance of the features. The fused mixed-scale feature pyramid can combine the advantages of features of different scales and improve the model's adaptability to scale changes and target detection performance.

[0064] In order to solve the problem that anchor-free detectors are greatly affected by scale changes in semi-supervised object detection, the present invention proposes a mixed-scale feature pyramid module. In semi-supervised object detection, scale change has always been a difficult challenge, especially for anchor-free detectors. Existing methods have many shortcomings in dealing with scale changes. The present invention constructs a mixed-scale feature pyramid to extract feature pyramids from conventional scales and downsampled views respectively. In terms of feature fusion, by adaptively fusing features of appropriate scales, the information of features of different scales is fully utilized.

[0065] Specifically, a feature fusion module is used to perform a squeeze operation (global average pooling to compress feature maps) followed by an excitation operation (learning the correlation between channels through two fully connected layers and generating weight values ​​for each channel), thereby achieving effective fusion of features of different scales. This design enables the model to better adapt to scale changes, improves the ability to detect objects of different sizes, and enhances the model's robustness to scale changes.

[0066] In summary, the present invention comprehensively solves the key problems faced by one-stage anchor-free detectors in semi-supervised target detection, such as low pseudo-label quality, complex post-processing, and poor scale adaptability, by introducing dense learning knowledge, a joint confidence estimation module, and a mixed-scale feature pyramid module. It significantly improves the performance of semi-supervised target detection and provides strong support for the promotion and development of target detection technology in practical applications.

[0067] Dense learning knowledge is introduced, and dense predictions are used to replace conventional sparse pseudo-boxes. Based on pixel-level predictions instead of unreliable pseudo-boxes, the continuous values ​​of the network output are directly used as a unified and direct form of pseudo-labels, eliminating the influence of post-processing hyperparameters on the model; a joint confidence estimation module is proposed to quantify the classification and positioning quality of pseudo-labels, and a region partitioning strategy is used to divide the learning area to generate more accurate pseudo-labels; a mixed-scale feature pyramid module is designed to adaptively fuse features of appropriate scales, indirectly improve the model through regularization, and enhance the model's robustness to scale changes.

[0068] Step 4: Use the generated mixed-scale feature pyramid to train the student model, optimize the model parameters, and finally obtain an optimized semi-supervised object detection model.

[0069] The training of the student model includes: using labeled data and generated pseudo labels as supervision information; optimizing model parameters through the weighted sum of supervised loss and unsupervised loss; and repeating the training process until the model converges.

[0070] Furthermore, the criteria for judging model convergence include but are not limited to: the performance of the model on the validation set reaches a preset threshold; the loss function of the model no longer decreases significantly within a certain number of iterations.

[0071] In addition, both the teacher model and the student model adopt a one-stage anchor-free detector architecture, including: the teacher model is used to generate high-quality dense pseudo labels; the student model is used to learn the pseudo labels generated by the teacher model.

[0072] The present invention significantly improves the quality of pseudo-labels by combining a confidence estimation module and a region partitioning strategy. In semi-supervised target detection, the accuracy of pseudo-labels is crucial to model performance. In traditional methods, pseudo-labels generated by a one-stage detector are often inaccurate, mainly because the classification score has a low correlation with the positioning quality. For example, in some complex scenarios, the target object may be occluded or deformed. It is difficult to accurately determine the true position and category of the target based solely on the classification score, resulting in low quality pseudo-labels generated by the teacher model, which in turn misleads the learning of the student model.

[0073] In summary, the joint confidence estimation module of the present invention adopts a dual-branch structure, one branch focuses on the classification task and outputs the classification score; the other branch is used to estimate the positioning quality and output the predicted IoU. By combining the two, a joint confidence is formed to more comprehensively evaluate the reliability of the pseudo-label. For example, in practical applications, for a partially occluded target, the classification branch may give a higher classification score, but the positioning branch will more accurately reflect the positioning quality of the target through the calculation of IoU based on the actual position and shape information of the target. Based on this joint confidence, regional division is performed, and the input image is divided into a learning area (including the top k% pixels with the highest score) and a suppression area (other areas are set to 0), and high-quality pseudo-labels are further screened out. In this way, the student model can obtain more accurate guidance during the learning process, reduce the misleading caused by low-quality pseudo-labels, and thus improve the learning efficiency and accuracy of the model.

[0074] The innovative approach of replacing pseudo-boxes with dense predictions greatly simplifies the model training process. Traditional pseudo-box-based methods need to go through a series of complex post-processing steps when generating pseudo-labels, such as non-maximum suppression (NMS), thresholding, and label assignment. These steps are not only computationally complex, but also introduce a large number of hyperparameters, such as NMS thresholds and score thresholds. In different application scenarios, these hyperparameters need to be fine-tuned, otherwise they will have a significant impact on model performance. For example, detecting pedestrians in monitoring scenarios and detecting product defects in industrial scenarios require different hyperparameter settings due to the different characteristics and distribution of the targets, which requires a lot of time and effort to adjust.

[0075] The dense prediction of the present invention directly uses the continuous values ​​of the network output as pseudo labels, cancels these post-processing steps, and fundamentally eliminates the influence of post-processing hyperparameters on the model. This not only reduces the amount of calculation in the training process, speeds up the training speed, but also makes the model training process more stable and reliable. At the same time, the dense pseudo-labels retain more information of the teacher model, provide richer learning resources for the student model, help the student model better understand the characteristics and distribution of the target, and further improve the performance of the model.

[0076] The mixed-scale feature pyramid module effectively enhances the model's ability to adapt to scale changes. In semi-supervised object detection, the scale variation of target objects is a common and critical challenge, especially for anchor-free detectors. Traditional methods are insufficient in dealing with scale changes. Although some existing multi-scale learning methods encourage the model to make scale-invariant predictions by adding downsampled views, they still rely on pseudo-labels generated from regular scales and cannot completely solve the problems caused by scale changes.

[0077] The hybrid scale feature pyramid of the present invention constructs a large-scale feature pyramid in a conventional view and a small-scale feature pyramid in a downsampled view respectively, and adaptively fuses the features of each scale through a feature fusion module. This multi-scale feature fusion method enables the model to better adapt to the detection requirements of targets of different scales, reduces the situation where the target is incorrectly assigned as the background in the pseudo-label due to scale changes, and improves the detection accuracy and robustness of the model for multi-scale targets in complex scenes.

[0078] In summary, the present invention comprehensively improves the performance of semi-supervised target detection by improving the quality of pseudo labels, simplifying the training process, enhancing scale adaptability and other innovations, providing strong support for the promotion and development of target detection technology in practical applications, and has important theoretical and practical significance.

[0079] Finally, it should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A lightweight semi-supervised target detection method based on joint estimation and scale fusion, characterized in that: The following steps are involved: Using labeled image data and unlabeled image data as input, a teacher model is used to generate dense pseudo labels for unlabeled images based on pixel-level predictions; Using a joint confidence estimation module to classify and estimate the positioning quality of the generated dense pseudo-labels, generate a joint confidence score, perform region division according to the joint confidence score, and divide the input image into a learning region and a suppression region; Construct a mixed-scale feature pyramid, including a large-scale feature pyramid in a conventional view and a small-scale feature pyramid in a downsampled view, and fuse the features of each scale through a feature fusion module to generate a mixed-scale feature pyramid; The generated mixed-scale feature pyramid is used for training the student model to optimize the model parameters, and finally an optimized semi-supervised target detection model is obtained.

2. According to claim 1, a lightweight semi-supervised target detection method based on joint estimation and scale fusion is characterized in that: The method generates dense pseudo labels for unlabeled images based on pixel-level predictions through a teacher model, including: Make pixel-level predictions for unlabeled images and assign target class probabilities to each pixel; Use the continuous values ​​output by the network as pseudo labels.

3. The lightweight semi-supervised target detection method based on joint estimation and scale fusion according to claim 1, characterized in that: The joint confidence estimation module is configured as a dual-branch structure, including: The classification branch is responsible for identifying the object category and outputting the classification score; The localization branch is used to estimate the localization quality and output the predicted intersection-over-union ratio; The outputs of the classification branch and the localization branch are combined through a sigmoid function to form a joint confidence score of the detection result.

4. The lightweight semi-supervised target detection method based on joint estimation and scale fusion according to claim 1, characterized in that: The performing of region division according to the joint confidence score includes: The input image is divided into a learning area and a suppression area. The learning area contains the first k% pixels with the highest scores, and the rest of the area is set to 0. The learned regions are used to generate high-quality pseudo labels, and the suppressed regions are used to reduce the interference of low-scoring predictions in background regions.

5. The lightweight semi-supervised target detection method based on joint estimation and scale fusion according to claim 1, characterized in that: The constructing of a mixed-scale feature pyramid includes: Constructing large-scale feature pyramids in regular views; Constructing small-scale feature pyramids in the downsampled views.

6. The lightweight semi-supervised target detection method based on joint estimation and scale fusion according to claim 1, characterized in that: The feature fusion module includes squeeze and excitation operations; The squeeze operation is used to perform global average pooling on the input feature map to generate channel descriptors; The excitation operation is used to learn the correlation between channels through two fully connected layers and generate weight values ​​for each channel. The weight values ​​are normalized by the Sigmoid activation function.

7. The lightweight semi-supervised target detection method based on joint estimation and scale fusion according to claim 1, characterized in that: The training of the student model includes: Use labeled data and generated pseudo labels as supervision information; Model parameter optimization is performed through the weighted sum of supervised loss and unsupervised loss; The training process is repeated until the model converges.

8. The lightweight semi-supervised target detection method based on joint estimation and scale fusion according to claim 7, characterized in that: The criteria for judging the convergence of the model include but are not limited to: The model's performance on the validation set reaches a preset threshold; The loss function of the model no longer decreases significantly within a certain number of iterations.

9. The lightweight semi-supervised target detection method based on joint estimation and scale fusion according to claim 1, characterized in that: Both the teacher model and the student model adopt a one-stage anchor-free detector architecture, including: The teacher model is used to generate high-quality dense pseudo labels; The student model is used to learn the pseudo labels generated by the teacher model.

10. The lightweight semi-supervised target detection method based on joint estimation and scale fusion according to claim 3, characterized in that: The classification branch adopts the FocalLoss focal loss function, and the positioning branch adopts the binary cross entropy loss function.

Images