CN-121999463-A - Small target traffic sign recognition method, medium and computer equipment

CN121999463ACN 121999463 ACN121999463 ACN 121999463ACN-121999463-A

Abstract

The invention relates to a small target traffic sign recognition method, medium and computer equipment, which take YOLOv as a basic model to construct an improved recognition model comprising a small target detection layer, construct a data set to train the improved recognition model, recognize the small target traffic sign by the trained improved recognition model, and realize the computer readable storage medium and the computer equipment based on the method. The method and the device better utilize shallow information of the small target traffic sign, reserve more small target traffic sign features, improve the positioning precision of the small target traffic sign, automatically pay attention to important feature channels, enhance the capturing capability of a network to key features, fuse semantic information and detail features more efficiently, improve the capturing capability of a model to feature graphic features, optimize the computing efficiency of feature extraction and fusion, effectively improve the feature utilization efficiency and the information richness, effectively enhance the trans-scale feature interaction capability of the model, and improve the detection capability of a detection head to the small target.

Inventors

YE LEI
LIANG DEYUAN
XU CHENQI
Wang Dihong

Assignees

浙江工业大学

Dates

Publication Date: 20260508
Application Date: 20251214

Claims (10)

1. A small target traffic sign recognition method is characterized in that the method takes YOLOv as a basic model to construct an improved recognition model, and the improved recognition model comprises a small target detection layer; and constructing a data set, training the improved recognition model, and recognizing the small target traffic sign by the trained improved recognition model.
2. The method for identifying small target traffic sign according to claim 1, wherein the improved identification model is based on YOLOv as a base model; replacing the C3K2 module by the enhanced C3K2 module through a backbone network of the model; The Neck part of the model comprises three groups of up-sampling layers, first full-connection layers, three groups of convolution modules and second full-connection layers which are sequentially arranged behind the backbone network, and the last two first full-connection layers are matched with the first two enhancement C3K2 modules of the backbone network; DCNv3-GSCSP modules are arranged behind each first full-connection layer and each second full-connection layer; the last four DCNv-GSCSP modules output through detection layers that introduce mixed attention HAttention modules, including small target detection layers.
3. The method for identifying small target traffic sign according to claim 1, wherein the enhanced C3K2 module comprises a convolution block, a segmentation block and a plurality of C3K modules decorated by SE, which are sequentially arranged, wherein the outputs of the segmentation block and the last C3K module decorated by SE are outputted after passing through the SDI feature fusion block and the convolution block.
4. The method for identifying small target traffic sign according to claim 3, wherein any one of the SE-modified C3K modules comprises a convolution layer and 2 Bottleneck _SE blocks connected in sequence, and outputs of the convolution layer and the Bottleneck _SE blocks are output through the convolution layer after being connected.
5. The method of claim 4, wherein Bottleneck _SE blocks comprise 2 convolution blocks, SE modules and fusion blocks connected in sequence.
6. The method for identifying the small target traffic sign according to claim 1, wherein the DCNv-GSCSP module comprises a first convolution layer, a GSConv branch and a second convolution layer are arranged behind the first convolution layer in parallel, outputs of the GSConv branch and the second convolution layer are added, and then a convolution result input with the DCNv3-GSCSP module is input into a full connection layer and a convolution layer which are sequentially connected and then output; the first convolution layer is a deformable convolution layer DCNv; the GSConv branches include 2 serially connected GSConv modules.
7. The method for identifying small target traffic sign according to claim 6, wherein the GSConv module comprises a convolution layer and a depth separable convolution layer which are sequentially arranged, and the outputs of the convolution layer and the depth separable convolution layer are output after passing through a full connection layer and a channel shuffling layer.
8. The method for identifying small target traffic sign according to claim 1, wherein the HAttention module comprises a shallow feature extraction unit and a deep feature extraction unit which are sequentially arranged, and outputs of the shallow feature extraction unit and the deep feature extraction unit are added and then processed and output through an image reconstruction unit.
9. A computer-readable storage medium, characterized in that a small target traffic sign recognition program is stored thereon, which program, when executed by a processor, implements the small target traffic sign recognition method according to one of claims 1 to 8.
10. A computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, characterized in that the processor implements the small target traffic sign recognition method according to one of claims 1 to 8 when executing the program.

Description

Small target traffic sign recognition method, medium and computer equipment Technical Field The invention relates to the technical field of image or video recognition or understanding, in particular to a small target traffic sign recognition method, medium and computer equipment. Background Object detection is one of the core tasks in the field of computer vision, aimed at identifying objects of interest in images or videos, and determining their location and size accurately. Unlike the mere image classification task, the object detection is not only required to judge the type of the object, but also required to solve the problem of positioning the object, and is considered as the basis of many other visual tasks, such as the need of first performing object detection to determine the position of the object in the case of example segmentation, the need of relying on object detection to identify the object to be marked in the image marking, the need of continuously monitoring the position of the detected object in the continuous frame in the object tracking, and the like. The application of the target detection covers a plurality of important directions such as pedestrian detection (used for identifying pedestrians in scenes such as intelligent security and automatic driving), face detection (widely applied to face recognition systems), text detection (playing a role in document analysis, image and text extraction and the like), traffic sign and traffic light detection (being a key link of an automatic driving and intelligent traffic system), remote sensing target detection (used in fields such as geographic information analysis and military reconnaissance) and the like. At the present time of rapid development of intelligent transportation systems (INTELLIGENT TRANSPORTATION SYSTEM, ITS) and automated driving technologies, traffic sign Recognition (TRAFFIC SIGN Recognition, TSR) has become a key research direction in the field of computer vision. The traffic sign is used as a core of road traffic management, and provides key information such as traffic rules, road conditions, safety prompts and the like for drivers and automatic driving systems. In an actual driving environment, however, the identification of traffic signs faces heavy challenges, including: (1) The direct irradiation of strong light in the daytime can cause reflection, and the dim light at night can cause the reduction of visibility, which can seriously interfere with clear imaging of traffic signs; (2) The traffic sign can be blurred in the rain and snow coverage, the visibility reduction in haze and the dust shielding in the dust and sand weather; (3) A shade, such as a roadside tree, building, or other vehicle, that may partially or completely shade a traffic sign; (4) The conditions of aging and fading, surface abrasion and the like can also reduce the discernability of the sign. The traditional traffic sign recognition method mainly relies on image processing technology and machine learning algorithm, such as edge detection, color segmentation, template matching and the like. Under a specific simple environment, the methods can play a certain role, but when facing complex scenes, higher recognition accuracy and robustness are often difficult to achieve. When the traffic sign is based on edge detection and recognition, if the traffic sign edge is unclear due to illumination or shielding, misjudgment is easy to occur, under the background of complex color, the recognition algorithm based on color segmentation is difficult to accurately segment the color of the traffic sign, the recognition algorithm based on template matching has strong dependence on templates, and once the traffic sign is deformed or angle-changed, the feature quantity is reduced, and the matching effect is greatly reduced. In recent years, the rise of the deep learning technology opens up a new path for a target detection task, wherein YOLO (You Only Look Once) series models are widely applied to the field of traffic sign recognition by virtue of high-efficiency, quick and accurate characteristics. However, the identification of the small target traffic sign is still a big problem in the field, because the small target traffic sign occupies very small proportion of pixels in the image and the characteristic information is deficient, so that the model is difficult to accurately extract the characteristics of the small target traffic sign, the small target is easy to be interfered by background noise, and the small target is more difficult to be effectively distinguished from the background under a complex background. Disclosure of Invention The invention solves the problems in the prior art, and provides a small-target traffic sign recognition method, medium and computer equipment, which are based on feature fusion and attention mechanism improvement YOLOv, so that the recognition accuracy of the small-target traffic sign is effectively improved. The technical scheme ado