CN-122023224-A - HD-YOLO model-based small target defect detection algorithm for high-performance polyethylene fiber surface

CN122023224ACN 122023224 ACN122023224 ACN 122023224ACN-122023224-A

Abstract

The invention discloses a small target defect detection algorithm of a high-performance polyethylene fiber surface based on an HD-YOLO model, which comprises the steps of acquiring images of the polyethylene fiber surface with defects by using image acquisition equipment, expanding a data set by adopting an image enhancement technology, labeling the enhanced data set, inputting the data set into the HD-YOLO network model for training, carrying out small target detection of the high-performance polyethylene fiber surface by using best. Pt of the trained HD-YOLO network model, designing a Hybrid Attentive Convolutional Fusion (HACF) module, and obviously improving the adaptability of the target detection model to the small target, a shielding target and a complex background on the premise of almost not increasing the calculation cost by using the HACF module, designing a Dual ATTENTIVE SPLIT Fusion (DASF) convolution module, thereby improving the problem of low accuracy in the small target defect detection process of the polyethylene fiber surface in the prior art, and having better production practicability.

Inventors

Zhang Shuaian
ZHOU XINJI
ZHU JIANJUN
NIU YANFENG
SHEN ZHIXIANG
CHEN JUNQI

Assignees

江苏九州星际新材料有限公司
浙江理工大学

Dates

Publication Date: 20260512
Application Date: 20250723

Claims (5)

1. The small target defect detection algorithm for the high-performance polyethylene fiber surface based on the HD-YOLO model is characterized by comprising the following steps of: S1, acquiring the surface of a high-performance polyethylene fiber with defects by using image acquisition equipment, and counting the number of various defects in a data set; s2, reconstructing the super-resolution of the image of the data set, and expanding the data set by adopting an image enhancement technology; S3, labeling the data set by utilizing MAKESENCE; s4, randomly dividing the data set into a training set, a verification set and a test set according to the proportion; s5, setting parameters batch_size, works and Epoch according to configuration information of a computer memory, a display card and a display memory which are actually operated; S6, loading weight of a pre-training model, and inputting a data set into the HD-YOLO model for training; The S7, YOLOv model mainly comprises an input end, a feature extraction network (Backbone), a neck network (Neck) and 4 parts of predictive output (Prediction), wherein the input end is used for preprocessing an image and inputting the image into the network, the Backbone part is used for extracting features, the Neck network part is used for further enhancing the diversity and the robustness of the extracted feature data, and the predictive part is used as an output end of the whole network and is responsible for outputting a target detection result; S8, adjusting the prior frame by a training prediction graph obtained by one round of network training to obtain a prediction frame, calculating IoU with a real frame of a target to represent the intersection ratio of the prediction frame and the real frame, obtaining positioning loss according to a target regression function CIoU, and weighting the classification loss, the confidence loss and the positioning loss to obtain the total loss of the network; s9, reversely transmitting the loss to the HD-YOLO model, and updating the network weight parameters by means of an SGD random gradient descent method to obtain a new weight model; S10, repeating the steps S6) -S9) through the Epoch value set in the step S5), wherein each round of updated network model parameters is used as a pre-training model of the next round to start a new round of training; S11, inputting images to be tested in a test set into an optimal model of S10), training, outputting the obtained prediction frames, restraining according to NMS maximum values, grouping all the prediction frames according to labels of each category, sorting the confidence degrees in the groups from large to small to obtain rectangular frames with highest scores, traversing the remaining rectangular frames, calculating the ratio of intersection and union of the rectangular frames with the highest scores, removing the remaining rectangular frames if the ratio is larger than a preset threshold, and continuously performing the operation on the remaining detection frames until the final prediction frames are obtained, thereby realizing small target defects on the surfaces of the high-performance polyethylene fibers.
2. The algorithm for detecting small target defects on the surface of the high-performance polyethylene fiber based on the HD-YOLO model according to claim 1, wherein the image enhancement technology in the step S2) expands the data set by normalizing the input image and enhancing the data set by adopting random gray scale, random brightness and contrast, random perspective transformation and random inversion methods.
3. The method for detecting small target defects on the surface of high-performance polyethylene fibers based on the HD-YOLO model according to claim 1, wherein the step S3) is characterized in that an obtained image is marked by using a wire mesh tool MAKESENCE, the type to be detected is preset, the small target defects on the surface of each type of polyethylene fibers to be detected are framed by using a rectangular frame, corresponding type labels are marked, TXT format of the YOLO labels is output after the labels are completed on all the images, the label content is the types of defects, and the coordinates (x, y) of the centers (x, y) of the rectangles and the width (width) and the length (height) of the rectangles are normalized.
4. The method for detecting small target defects on the surface of high-performance polyethylene fiber based on HD-YOLO model according to claim 1, wherein the HD-YOLO model in step S6) comprises adding a median filter at the input end of YOLOv and adding a designed Hybrid Attentive Convolutional Fusion module (HACF module) at the trunk part of YOLOv11, and replacing the conventional convolution with a Dual ATTENTIVE SPLIT Fusion convolution module (DASF module), wherein A. The median filter is used for selecting pixel values of a digital image or a digital sequence and the pixel values of adjacent pixel points around the pixel points, sequencing the pixel values, taking the pixel value positioned in the middle position as the pixel value of the current pixel point, and enabling the pixel value around to be close to a true value, so that isolated noise points are eliminated; the HACF module remarkably improves the adaptability of the target detection model to small targets, shielding targets and complex backgrounds on the premise of hardly increasing the calculation cost through the attention and convolution mixed architecture and the dynamic characteristic fusion mechanism; and c, the DASF module is cooperatively enhanced and dynamically guided by the multi-domain features, and the DASF module remarkably improves the robustness of the small target detection model in a complex scene, and meanwhile, the higher calculation efficiency is maintained.
5. The algorithm for detecting small target defects on the surface of the high-performance polyethylene fiber based on the HD-YOLO model according to claim 1, wherein the step S8) is as follows: a. the prior frame is adjusted by the training prediction graph to obtain a prediction frame, then the prediction frame and a real frame of a target are calculated IoU to represent the ratio of the intersection of the prediction frame and the real frame to the union, and then the positioning loss is obtained according to the target regression function CIoU: b. The classification loss is calculated using a binary cross entropy function: c. As with the calculation loss approach described above, the confidence loss L obj is calculated using a binary cross entropy function, and then the classification loss, confidence loss, location loss are weighted to obtain the total loss of the network: LOSS=λ 1 L cls +λ 2 L obj +λ 3 L loc 。

Description

HD-YOLO model-based small target defect detection algorithm for high-performance polyethylene fiber surface Technical Field The invention relates to the field of intelligent image processing and machine vision, in particular to a high-performance polyethylene fiber defect detection algorithm based on an HD-YOLO model according to claim 1. Background High performance polyethylene fibers are increasingly used in industrial applications, but are prone to structural defects during their manufacture. The traditional manual detection method is difficult to adapt to the continuous high-speed running rhythm of an automatic production line, so that the detection efficiency and the production speed are seriously disjointed. The detection personnel face high repeatability work for a long time and are easy to generate visual fatigue and distraction, so that the omission ratio of defects such as micron-sized fiber fracture, abnormal crystallinity and the like is very high, and the tensile strength and corrosion resistance of the product are directly affected. In addition, precision laboratory detection equipment is purchased at a cost exceeding a million-scale, which makes medium and small-sized manufacturing enterprises face heavy quality control pressures. The traditional industrial detection system relies on high-end equipment such as a precise spectrometer, an electron microscope and the like to construct a multi-mode detection system, and a professional optical coupling device, a high-precision mechanical transmission platform and a complex signal processing algorithm are required to be configured, so that the system integration level is low and the maintenance cost is high. Although the current YOLOv and other deep learning models of the mainstream show remarkable advantages in the field of target detection, when the defects of the nano-scale structures in the fiber are faced, the detection effect is not ideal due to small and complex characteristic dimensions. Disclosure of Invention In view of the above background, in order to improve the defect detection accuracy of the high-performance polyethylene fiber, a high-performance polyethylene fiber defect detection algorithm based on an HD-YOLO model is provided. Firstly, aiming at the problem of noise pollution in an industrial detection scene, a median filter is added at the input end of YOLOv, the detection performance is indirectly improved by inhibiting spiced salt noise and optimizing the quality of a feature map, the detection precision of a model on defects in a small target and a complex environment is improved, secondly, aiming at the problems of insufficient multi-scale feature Fusion capability and limited small target detection precision of YOLOv11 in the complex industrial scene, a Hybrid Attentive-Convolutional Fusion (HACF) module is designed, the HACF module remarkably improves the adaptability of a target detection model to the small target, a shielding target and the complex background on the premise of almost not increasing the calculation cost by a concentration-convolution mixed architecture and a dynamic feature Fusion mechanism, and finally, aiming at the problem of serious fine-grained feature loss of the small target defect in the traditional convolution downsampling process, a Dual ATTENTIVE SPLIT Fusion (DASF) convolution module is designed, and the robustness of the target detection model under the complex scene is remarkably improved by a multi-domain feature enhancement and dynamic attention guidance, and meanwhile, the higher calculation efficiency is kept. The technical scheme adopted by the invention is as follows: A high-performance polyethylene fiber defect detection algorithm based on an HD-YOLO model comprises the following steps: S1, acquiring high-performance polyethylene fibers with defects by using image acquisition equipment, and counting the number of various defects in a data set; s2, reconstructing the super-resolution of the image of the data set, and expanding the data set by adopting an image enhancement technology; S3, labeling the data set by utilizing MAKESENCE; s4, randomly dividing the data set into a training set, a verification set and a test set according to the proportion; s5, setting parameters batch_size, works and Epoch according to configuration information of a computer memory, a display card and a display memory which are actually operated; S6, loading weight of a pre-training model, and inputting a data set into the HD-YOLO model for training; The S7, YOLOv model mainly comprises an input end, a feature extraction network (Backbone), a neck network (Neck) and 4 parts of predictive output (Prediction), wherein the input end is used for preprocessing an image and inputting the image into the network, the Backbone part is used for extracting features, the Neck network part is used for further enhancing the diversity and the robustness of the extracted feature data, and the predictive part is used as an ou