Device, method and equipment for detecting and classifying small target images
Through the sharing feature extraction module and the method of cutting local feature blocks, the existing small object detection methods are solved, and more efficient computing resource utilization and model training simplification is achieved.
Patent Information
- Application Number
- CN202411492530.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-24
- Publication Date
- 2025-02-18
AI Technical Summary
The existing small object detection methods have problems such as large calculations, long time, and complex data production and model training, especially when small objects are easily obstructed and their features are not obvious enough.
By sharing the feature extraction module, avoiding repeated calculations of the same features, thereby reducing the amount of calculation, VGG or FPN feature extraction network is used to extract image features, and local feature blocks are cropped on the feature map for subsequent processing.
It reduces the waste of computing resources, improves the utilization rate of computing resources, simplifies model training and deployment, and improves inference speed and model accuracy.
Smart Images
Figure CN119478346A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of small target detection, and in particular to a device, method and equipment for small target image detection and classification. Background Art
[0002] Small targets usually refer to targets in the distant small target scene, where the target size in the image is small and it is easy to be misidentified or ignored. Small target detection is an important topic in the field of computer vision, especially in monitoring, unmanned driving, medical imaging and other applications. Taking personnel security control as an example, in the detection, it is not only necessary to detect the person, but also to judge some other characteristics of the person, such as clothing, gender, age, backpack and other auxiliary features. The existing conventional technology is divided into two consecutive steps. First, the person is detected, and then the auxiliary characteristics of the person are obtained by detection segmentation or classification based on the cut-out person image. This method consumes more time in data production and model training.
[0003] Therefore, existing small target detection methods generally have the following problems and defects: 1. Small targets are easily obscured by the background or other objects due to their small scale and only occupy a small area; 2. Small targets have fewer pixels, so their features are not obvious enough and it is difficult for the detector to directly extract useful information; 3. Although the use of multi-stage detection has improved the effect of small target detection to a certain extent, it increases the time consumption and the time consumption of early data production, model training, and later deployment of multiple models. Summary of the invention
[0004] In view of the defects in the prior art, the purpose of the present invention is to provide a device, method and equipment for small target image detection and classification, which avoids repeated calculation of the same features by sharing the feature extraction module, thereby reducing the amount of calculation and improving the utilization of computing resources.
[0005] To solve the above problems, the technical solution of the present invention is:
[0006] A device for small target image detection and classification comprises an image input module, a feature extraction module and a detection or classification module; the image input module is used to process the collected image and input it into the feature extraction module; the feature extraction module is used to extract the features of the image through a feature extraction model to obtain the coordinate information of the small target; the detection or classification module is used to cut out a local feature block on a feature map according to the coordinate information of the small target and input it into a decision head to obtain the feature information of all small targets in the image.
[0007] Preferably, the feature extraction module uses a VGG feature extraction network or an FPN feature extraction network to extract image features.
[0008] Preferably, the cropped local feature block only contains the required target area, thereby reducing noise in model learning.
[0009] Furthermore, the present invention also provides a method for detecting and classifying small target images, comprising the following steps:
[0010] The collected images are processed through the image input module and input into the feature extraction model;
[0011] The feature extraction model is used to extract features from the input image to obtain the coordinate information of the small target;
[0012] According to the coordinate information of the small targets, local feature blocks are cropped from the feature map and input into the decision head to obtain the feature information of all small targets in the image.
[0013] Furthermore, the present invention also provides a computer device, comprising a processor and a memory for storing executable instructions of the processor, wherein the processor is configured to execute the above-mentioned method for small object image detection and classification by executing the executable instructions.
[0014] Compared with the prior art, the present invention has the following beneficial effects:
[0015] 1. The method of the present invention avoids repeated calculation of the same features by sharing the feature extraction part, thereby reducing the amount of calculation and improving the utilization rate of computing resources.
[0016] 2. The present invention modifies the conventional target detection algorithm model and adds a classification head for small target classification or a detection head for small target detection after feature extraction, thereby simplifying model training and deployment. The merged model only needs to be trained once, avoiding the complexity brought by the training of multiple-stage models separately.
[0017] 3. The method of the present invention unifies the optimization objectives. In the merged model, each stage can share the optimization objectives. Through joint training, the model can better coordinate the tasks of each stage to achieve better overall performance. Moreover, the parameters between the stages can influence each other. Through joint optimization, the parameters of each stage can be better adjusted to make the model more stable and efficient. BRIEF DESCRIPTION OF THE DRAWINGS
[0018] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:
[0019] Figure 1 It is a schematic diagram of the structure of a conventional device for small target image detection and classification;
[0020] Figure 2 It is a schematic diagram of the structure of the device for small object image detection and classification of the present invention;
[0021] Figure 3 The figure is a flowchart of the method for detecting and classifying small target images according to the present invention. DETAILED DESCRIPTION
[0022] The present invention is described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, several changes and improvements can also be made without departing from the concept of the present invention. These all belong to the protection scope of the present invention.
[0023] Specifically, the present invention provides a device for detecting and classifying small target images, such as Figure 1 and Figure 2 As shown, the device for small target image detection and classification includes an image input module, a feature extraction module and a detection or classification module. The image input module is used to process the collected image and input it into the feature extraction module. The feature extraction module is used to extract the features of the image through the model to obtain the coordinate information of the small target. The feature extraction module uses the existing VGG feature extraction network or FPN feature extraction network to extract the image features.
[0024] The detection or classification module is used to cut out local feature blocks on the feature map according to the coordinate information of the small target and input them into the decision head to obtain the feature information of all small targets in the image. The present invention modifies the conventional target detection algorithm model, adds a classification head for small target classification or a detection head for small target detection after extracting features, and uses the same feature extraction structure to perform subsequent personnel detection and detection or classification of personnel's ancillary features. All information on the map can be obtained through one reasoning, which plays a role in reducing the model time consumption and synchronously obtaining ancillary feature information.
[0025] The conventional small target detection and classification model is constructed as follows Figure 1 As shown in the figure, a multi-level deep model is used to input the original image. After extracting the coordinate information of the small target, the small target image is cropped from the original image and then the feature is extracted according to the next model to extract the information of the small target. Taking the feature information of a person as an example, under normal circumstances, it is necessary to first detect the coordinates of the person through the person detection model, then cut out the person image according to the coordinates, and then pass it into the new model to obtain the characteristics of the person and other information.
[0026] like Figure 2As shown, compared with the conventional technical solution, the method of the present invention omits the small target feature extraction structure and the steps of cutting small images on the original image according to the intermediate results and then entering the new feature extraction module. The core of the present invention is to share the same feature extraction module. The later module directly cuts out the feature information map according to the small target coordinate value obtained in the early stage of detection as the later input. In this way, the computational redundancy is reduced by sharing the feature extraction module, especially in models with high computational load, which can save more computing resources and is particularly suitable for scenarios with limited computing resources such as embedded or edge devices. Moreover, when extracting features, global context information is also reflected. The global features can be fused with the attention mechanism before cutting to help subsequent modules obtain richer information. It has more features than the conventional multi-level directory cutting image and then extracting more features, and can also combine the feature information of objects around people.
[0027] The advantages of the present invention are as follows: the reasoning speed is improved, the feature extraction module is executed only once, and the subsequent clipping and further processing depend on the previous detection results, which simplifies the input of the later model, thus reducing the amount of data required to be processed in the subsequent steps, thereby improving the reasoning speed; the model is highly modular and reusable, and this modular design facilitates model upgrades and maintenance. For example, the feature extraction module can be independently optimized or replaced without affecting the later modules.
[0028] Data cropping focuses on key areas. Cropping the person feature information map as input can help the later model focus on key areas, reduce interference information, and improve the accuracy of the model. For example, in subsequent human posture recognition or behavior analysis tasks, the cropped image only contains the required target area, which can reduce noise in model learning.
[0029] Furthermore, the present invention also provides a method for detecting and classifying small target images, such as Figure 3 As shown, the method comprises the following steps:
[0030] S1: Process the collected image through the image input module and input it into the feature extraction model;
[0031] S2: Extract features from the input image through the feature extraction model to obtain the coordinate information of the small target;
[0032] S3: According to the coordinate information of the small target, the local feature block is cut out from the feature map and input into the decision head to obtain the feature information of all small targets in the image.
[0033] Specifically, the present invention modifies the conventional target detection algorithm model, shares the same feature extraction module, and cuts out a feature information map as a later input based on the small target coordinate values obtained in the early stage of detection. This reduces computational redundancy by sharing the feature extraction module, especially in models with high computational complexity, and can save computing resources.
[0034] According to another aspect of an embodiment of the present application, a computer device is provided, which includes a processor and a memory for storing executable instructions of the processor, and the processor is configured to execute the method for small object image detection and classification described in the embodiment of the present application by executing the executable instructions.
[0035] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essence of the present invention. In the absence of conflict, the embodiments of the present application and the features in the embodiments can be combined with each other arbitrarily.
Claims
1. A device for detecting and classifying small target images, characterized in that: The device comprises an image input module, a feature extraction module and a detection or classification module; the image input module is used to process the collected image and input it into the feature extraction module; The feature extraction module is used to extract the features of the image through a feature extraction model to obtain the coordinate information of the small target; the detection or classification module is used to crop the local feature blocks on the feature map according to the coordinate information of the small target and input them into the decision head to obtain the feature information of all small targets in the image.
2. The device for detecting and classifying small target images according to claim 1, characterized in that: The feature extraction module uses a VGG feature extraction network or an FPN feature extraction network to extract image features.
3. The device for detecting and classifying small target images according to claim 1, characterized in that: The cropped local feature block only contains the required target area, thereby reducing noise in model learning.
4. A method for detecting and classifying small target images, characterized in that: The method comprises the following steps: The collected images are processed through the image input module and input into the feature extraction model; The feature extraction model is used to extract features from the input image to obtain the coordinate information of the small target; According to the coordinate information of the small targets, local feature blocks are cropped from the feature map and input into the decision head to obtain the feature information of all small targets in the image.
5. The method for detecting and classifying small target images according to claim 4, characterized in that: The feature extraction model uses a VGG feature extraction network or an FPN feature extraction network to extract image features.
6. The method for detecting and classifying small target images according to claim 4, characterized in that: The cropped local feature block only contains the required target area, thereby reducing noise in model learning.
7. A computer device, characterized in that: The computer device includes a processor and a memory for storing executable instructions of the processor, and the processor is configured to perform the method for small object image detection and classification according to any one of claims 4 to 6 by executing the executable instructions.