Search

CN-121981944-A - Clothing flaw detection method and system based on multi-spectral feature harmonic aggregation complex shape

CN121981944ACN 121981944 ACN121981944 ACN 121981944ACN-121981944-A

Abstract

The invention provides a clothing flaw detection method and a clothing flaw detection system based on multi-frequency spectrum feature harmonic aggregation, which relate to the technical field of image detection, wherein a multi-frequency spectrum feature harmonic aggregation flaw detection model is established and pre-trained, the multi-frequency spectrum feature harmonic aggregation flaw detection model comprises a frequency spectrum domain frequency division network, a frequency spectrum harmonic aggregation network, a side spectrum complex network and a target detection output layer, an image to be detected in original clothing image data is input into the multi-frequency spectrum feature harmonic aggregation flaw detection model, the image to be detected is sequentially processed through the frequency spectrum domain frequency division network, the frequency spectrum harmonic aggregation network and the side spectrum complex network, a feature representation with enhanced edge perception capability is obtained, the target detection and classification are carried out on the feature representation through the target detection layer, and a flaw detection result of the image to be detected is output. The invention solves the problems of difficult identification of the clothes flaws, low manual detection efficiency and the like in the actual production process, and realizes efficient and accurate detection of the clothes flaws in the application scene of the actual engineering.

Inventors

  • XU LIANG
  • ZHU GUANHENG

Assignees

  • 广东工业大学

Dates

Publication Date
20260505
Application Date
20251211

Claims (10)

  1. 1. The clothing flaw detection method based on multi-spectral characteristic harmonic aggregation complex shape is characterized by comprising the following steps of: acquiring original clothing image data, and preprocessing the acquired original clothing image data to obtain a clothing image to be detected; Establishing a multi-frequency spectrum characteristic harmonic aggregation complex flaw detection model and performing pre-training, wherein the multi-frequency spectrum characteristic harmonic aggregation complex flaw detection model comprises a spectrum domain frequency division network layer, a frequency spectrum harmonic aggregation network layer, a side spectrum complex network layer and a target detection layer which are connected in sequence; Carrying out frequency domain decomposition and energy reconstruction on deep features of the clothing image to be detected by utilizing the spectral domain frequency division network layer, and outputting multi-layer spectrum characterization; utilizing the spectrum harmonic aggregation network layer to carry out resonance fusion and semantic enhancement on the multi-layer spectrum characterization and outputting high-resolution aggregation characteristics; Performing self-adaptive multiplexing and structure enhancement on the feature boundary of the high-resolution aggregation feature by utilizing a side spectrum multiplexing network layer to generate a feature representation with enhanced edge perception capability; And outputting a clothing flaw detection result through the target detection output layer based on the characteristic representation with the enhanced edge perception capability.
  2. 2. The method for detecting the flaws in the clothing based on the harmonic aggregation of the multi-spectral features according to claim 1, wherein the preprocessing comprises the following steps: labeling the flaws in the original clothing image data to obtain a target detection data set; And performing random overturning, rotation, shearing transformation and mixed enhancement operation on the target image data in the target detection data set, and performing normalization and scaling to obtain the clothing image to be detected.
  3. 3. The method for detecting flaws in clothing based on multi-spectral feature harmonic complexes of claim 1 wherein the spectral domain division network layer comprises a frequency filtering layer, a learnable frequency response layer and a frequency band decomposition layer.
  4. 4. The method for detecting the flaws of the clothing based on the harmonic aggregation complex shape of the multi-spectral features according to claim 3, wherein the process of performing frequency domain decomposition and energy reconstruction is as follows: inputting the clothing image to be detected into the frequency filtering layer, carrying out frequency analysis on the image characteristics of the clothing image to be detected, and extracting spectral energy distribution information; Performing weighted modulation on different frequency components of an image by using the learnable frequency response layer, and adaptively adjusting frequency response weights according to frequency domain energy intensity in the frequency spectrum energy distribution information to generate frequency spectrum modulation characteristics; And inputting the frequency spectrum modulation characteristics into the frequency band decomposition layer, dividing the frequency spectrum modulation characteristics into a high-frequency component and a low-frequency component according to a preset frequency threshold, and respectively outputting the high-frequency component and the low-frequency component to obtain a multi-frequency characteristic sub-band to obtain a multi-layer frequency spectrum characterization.
  5. 5. The method for detecting flaws in clothing based on multi-spectral feature harmonic aggregation as claimed in claim 4, wherein the spectral harmonic aggregation network layer comprises a frequency attention layer, a guide aggregation layer and a feature fusion layer.
  6. 6. The method for detecting the clothing flaws based on the multi-frequency characteristic harmonic aggregation complex form according to claim 5, wherein the process of carrying out resonance fusion and semantic reinforcement is as follows: Inputting the high-frequency component into the frequency attention layer, and performing response amplification on the high-frequency component in the frequency attention layer by using an attention mechanism to obtain high-frequency attention characteristics; Based on the high-frequency attention characteristic and the low-frequency attention characteristic, carrying out weighted linear fusion on the multi-layer spectrum characterization by utilizing the guide aggregation layer to generate a harmonic aggregation fusion characteristic; And inputting the harmonic aggregation fusion features into the feature fusion layer, and outputting high-resolution aggregation features in the feature fusion layer through a multi-scale convolution and cross-layer connection mechanism.
  7. 7. The method for detecting the flaws of the clothing based on the multi-spectral feature harmonic aggregation complex of claim 1, wherein the edge spectral complex network layer comprises an edge perception layer, a frequency spectrum feedback layer and a structural complex enhancement layer.
  8. 8. The method for detecting the flaws in the clothing based on the multi-frequency characteristic harmonic aggregation complex form according to claim 7, wherein the process of performing the self-adaptive complex form and the structure enhancement is as follows: Inputting the high-resolution aggregation features into the edge perception layer, extracting boundary information of the high-resolution aggregation features, and obtaining an initial edge perception graph; through the frequency spectrum feedback layer, characteristic mutual feedback is carried out between a space domain and a frequency domain, edge high-frequency components and semantic low-frequency components in the initial edge perception diagram are subjected to bidirectional interaction, and a boundary blurring and detail breaking area is repaired; And carrying out multi-step recursive updating on the high-resolution aggregation features through a learnable weighted gating mechanism by utilizing the structural complex enhancement layer, and generating a feature representation with enhanced edge perception capability.
  9. 9. The clothing flaw detection method based on multi-spectral feature harmonic aggregation complex of claim 1, wherein the target detection layer comprises a feature prediction sub-layer, a bounding box regression sub-layer and a category judgment sub-layer, and the processing procedure of the target detection layer comprises: Inputting the characteristic representation with the enhanced edge perception capability into the characteristic prediction sublayer, extracting response characteristics of a candidate target area through a multi-scale convolution structure, and generating a target candidate characteristic diagram; inputting the target candidate feature map into the boundary box regression sub-layer, carrying out regression prediction on the space position and the size of the candidate target region based on an anchor box mechanism, and determining the boundary box coordinates of the flaw region; Performing feature classification on the candidate target area by using the class judgment sub-layer, calculating confidence coefficient distribution of each flaw class, and outputting probability prediction results of each flaw class; Performing non-maximum value inhibition treatment on the boundary frame coordinates of the flaw area and the confidence coefficient of the flaw category, and marking the position coordinates of the flaw area by using the processed boundary frame coordinates of the flaw area; and outputting a flaw detection result comprising the position coordinates of the flaw area, the flaw category and the confidence of the flaw category.
  10. 10. A clothing flaw detection system based on multi-spectral feature harmonic aggregation complex shape for realizing the clothing flaw detection method based on multi-spectral feature harmonic aggregation complex shape according to claim 1, comprising: the image acquisition preprocessing module is used for acquiring original clothing image data, preprocessing the acquired original clothing image data and obtaining a clothing image to be detected; The model construction module is used for building a multi-frequency spectrum characteristic harmonic aggregation complex flaw detection model and pre-training, and the multi-frequency spectrum characteristic harmonic aggregation complex flaw detection model comprises a spectrum domain frequency division network layer, a frequency spectrum harmonic aggregation network layer, a side spectrum complex network layer and a target detection layer which are connected in sequence; The characteristic decomposition module is used for carrying out frequency domain decomposition and energy reconstruction on the deep characteristic of the clothing image to be detected by utilizing the spectral domain frequency division network layer and outputting multi-layer spectrum characterization; the characteristic aggregation module is used for carrying out resonance fusion and semantic reinforcement on the multi-layer spectrum characterization through the spectrum harmonic aggregation network layer and outputting high-resolution aggregation characteristics; the edge reconstruction module is used for carrying out self-adaptive shaping and structure enhancement on the feature boundary of the high-resolution aggregation feature by utilizing the side spectrum shaping network layer to generate a feature representation with enhanced edge perception capability; And the flaw detection module is used for carrying out target detection and classification on the characteristic representation with the enhanced edge perception capability by utilizing the target detection layer and outputting a flaw detection result of the clothing image to be detected.

Description

Clothing flaw detection method and system based on multi-spectral feature harmonic aggregation complex shape Technical Field The invention relates to the technical field of image detection, in particular to a clothing flaw detection method and system based on multi-spectral characteristic harmonic aggregation complex shape. Background The textile industry is the prop industry in China, the problem fabric with flaws is inevitably produced in the textile production process, and the flaw detection of clothes is one of the indispensable links in the production process. The conventional method is excellent in terms of feature recognition accuracy and detection speed, however, there are some limitations to flaw detection in some cases. For example, conventional methods generally deal with limited types of flaws, and are less accurate in handling small flaws in dark garments, thread ends, etc. In addition, the stability is low when flaw detection is performed on complex printed garments. For defects with complex texture background and large scale range, the detection effect of the traditional method still has room for further improvement. Moreover, the conventional method cannot realize timely, high-speed and accurate real-time detection. The machine vision flaw detection algorithm has made a certain progress in replacing manual detection, but the existing network model has higher computational complexity, higher input cost and limited range of problems to be solved. With the continuous development of the textile industry, the variety and process steps of the clothing are increasing, and the number of defects of textile products is characterized by numerous, different sizes, tiny blemishes or high concealment. However, the existing flaw detection technology in the textile industry has a limited detection range, and cannot clearly, efficiently and accurately identify all flaws. Disclosure of Invention In order to solve the problem of low accuracy in detection of clothes flaws in the prior art, the invention provides a clothes flaw detection method and a clothes flaw detection system based on multi-frequency characteristic harmonic aggregation complex shape, which realize efficient and accurate detection of clothes flaws in a realistic engineering application scene. In order to achieve the technical effects, the technical scheme of the invention is as follows: A clothing flaw detection method based on multi-spectral characteristic harmonic aggregation complex shape comprises the following steps: acquiring original clothing image data, and preprocessing the acquired original clothing image data to obtain a clothing image to be detected; Establishing a multi-frequency spectrum characteristic harmonic aggregation complex flaw detection model and performing pre-training, wherein the multi-frequency spectrum characteristic harmonic aggregation complex flaw detection model comprises a spectrum domain frequency division network layer, a frequency spectrum harmonic aggregation network layer, a side spectrum complex network layer and a target detection layer which are connected in sequence; Carrying out frequency domain decomposition and energy reconstruction on deep features of the clothing image to be detected by utilizing the spectral domain frequency division network layer, and outputting multi-layer spectrum characterization; utilizing the spectrum harmonic aggregation network layer to carry out resonance fusion and semantic enhancement on the multi-layer spectrum characterization and outputting high-resolution aggregation characteristics; Performing self-adaptive multiplexing and structure enhancement on the feature boundary of the high-resolution aggregation feature by utilizing a side spectrum multiplexing network layer to generate a feature representation with enhanced edge perception capability; And outputting a clothing flaw detection result through the target detection output layer based on the characteristic representation with the enhanced edge perception capability. Preferably, the pretreatment includes: labeling the flaws in the original clothing image data to obtain a target detection data set; And performing random overturning, rotation, shearing transformation and mixed enhancement operation on the target image data in the target detection data set, and performing normalization and scaling to obtain the clothing image to be detected. Preferably, the spectral domain division network layer includes a frequency filtering layer, a learnable frequency response layer, and a band decomposition layer. Preferably, the frequency domain decomposition and energy reconstruction process is as follows: inputting the clothing image to be detected into the frequency filtering layer, carrying out frequency analysis on the image characteristics of the clothing image to be detected, and extracting spectral energy distribution information; Performing weighted modulation on different frequency components of an image by using th