CN-121982007-A - Defect detection method, device, equipment and medium for medical consumables

CN121982007ACN 121982007 ACN121982007 ACN 121982007ACN-121982007-A

Abstract

The invention discloses a defect detection method, device, equipment and medium for medical consumables. The method comprises the steps of obtaining multi-mode data of nondefective medical consumables as normal sample data, enabling the multi-mode data to comprise visible light images, infrared images and three-dimensional point cloud information, adding interference to the normal sample data to generate pseudo-defect sample data, determining multi-mode training data pairs according to the normal sample data and the pseudo-defect sample data, extracting features of the multi-mode training data pairs to obtain multi-mode feature information, carrying out feature alignment on the multi-mode feature information to obtain target feature information through contrast learning, inputting the target feature information into a defect detection model to carry out self-supervision training so as to detect defects of the medical consumables to be detected by using the trained defect detection model, and carrying out defect detection on the defect detection model based on differences between the target feature information and reconstruction results of the target feature information. The scheme can expand the defect detection coverage while reducing defect labeling dependence.

Inventors

Xiong Kezan
LU ZHEMING
WenRen Ji

Assignees

苏州桓球医疗科技有限公司

Dates

Publication Date: 20260505
Application Date: 20260129

Claims (10)

1. A medical consumable-oriented defect detection method, the method comprising: acquiring multi-mode data of defect-free medical consumables as normal sample data, wherein the multi-mode data comprises visible light images, infrared images and three-dimensional point cloud information; adding interference to the normal sample data to generate pseudo-defect sample data, and determining multi-mode training data pairs according to the normal sample data and the pseudo-defect sample data; extracting features of the multi-modal training data pairs to obtain multi-modal feature information, and carrying out feature alignment on the multi-modal feature information through contrast learning to obtain target feature information; and inputting the target characteristic information into a defect detection model for self-supervision training so as to detect the defects of the medical consumable to be detected by using the trained defect detection model, wherein the defect detection model detects the defects based on the difference value between the target characteristic information and the reconstruction result.
2. The method of claim 1, wherein adding interference to the normal sample data generates pseudo-defect sample data, comprising: Generating a defect mask in the visible light image and implanting texture disturbance to obtain pseudo defect sample data corresponding to the visible light image; Generating radiation intensity interference in the infrared image to obtain pseudo defect sample data corresponding to the infrared image; And adding deformation noise to the local area point cloud in the three-dimensional point cloud information to obtain pseudo defect sample data corresponding to the three-dimensional point cloud information.
3. The method according to claim 1, wherein the performing feature alignment on the multi-modal feature information through contrast learning to obtain target feature information includes: Mapping the multi-modal feature information to a target feature space through nonlinear mapping; and performing self-supervision contrast learning on the multi-mode feature information of the target feature space based on the contrast loss function to obtain the target feature information.
4. A method according to claim 3, wherein the contrast loss function comprises a local contrast loss function for measuring feature alignment effects of two modal feature information and a global contrast loss function for measuring feature alignment effects of multi-modal feature information, the global contrast loss function being constructed based on the local contrast loss function.
5. The method of claim 1, wherein inputting the target feature information into a defect detection model for self-supervised training comprises: Inputting the target characteristic information into a reconstruction network in a defect detection model, carrying out data reconstruction on the target characteristic information through the reconstruction network to obtain a reconstruction result, and determining a multi-mode reconstruction error according to a difference value between the target characteristic information and the reconstruction result; And inputting the multi-modal reconstruction errors to a multi-modal judging module in a defect detection model, carrying out self-adaptive weighted fusion on the multi-modal reconstruction errors through the multi-modal judging module to obtain a fusion abnormal result, and determining a defect segmentation result and a defect score based on the fusion abnormal result.
6. The method of claim 5, wherein the reconstruction network employs a U-shaped network structure, the reconstruction network comprises an encoder and a decoder, the encoder is constructed based on a depth separable convolution and an adaptive discrete wavelet transform, and the multi-modal discrimination module comprises an attention mechanism fusion module and a discriminator.
7. The method according to any one of claims 1-6, further comprising, after acquiring the multi-modal data of the defect-free medical consumable as normal sample data: And preprocessing the normal sample data, wherein the preprocessing comprises denoising enhancement, spatial registration, size alignment and data division.
8. A medical consumable-oriented defect detection device, the device comprising: the normal sample acquisition module is used for acquiring multi-mode data of the medical consumable without defects as normal sample data, wherein the multi-mode data comprises visible light images, infrared images and three-dimensional point cloud information; The pseudo-defect sample construction module is used for adding interference to the normal sample data to generate pseudo-defect sample data, and determining multi-mode training data pairs according to the normal sample data and the pseudo-defect sample data; the feature alignment module is used for extracting features of the multi-mode training data pairs to obtain multi-mode feature information, and carrying out feature alignment on the multi-mode feature information through contrast learning to obtain target feature information; the model training module is used for inputting the target characteristic information into a defect detection model for self-supervision training so as to detect defects of the medical consumable to be detected by using the trained defect detection model, wherein the defect detection model detects the defects based on the difference value between the target characteristic information and the reconstruction result of the target characteristic information.
9. An electronic device, the electronic device comprising: at least one processor, and A memory communicatively coupled to the at least one processor, wherein, The memory stores a computer program executable by the at least one processor to enable the at least one processor to perform the medical consumable-oriented defect detection method of any one of claims 1-7.
10. A computer readable storage medium, characterized in that it stores computer instructions for causing a processor to implement the medical consumable oriented defect detection method of any of claims 1-7 when executed.

Description

Defect detection method, device, equipment and medium for medical consumables Technical Field The invention relates to the technical field of computer vision, in particular to a defect detection method, device, equipment and medium for medical consumables. Background Medical consumables generally have the characteristics of mass production, relatively low single-piece value, important safety responsibility, various defect types, low defect sample ratio and the like. During the production, packaging, transportation and storage of medical consumables, defects such as cracks, gaps, depressions, flaking, abrasion, corrosion, deformation, abnormal surface texture and the like may occur. If the defect is not detected in time, there may be a risk of clinical use and a risk of quality traceability. Traditional detection relies on manual visual inspection or single-mode sensor detection. Manual visual inspection is influenced by subjective experience, illumination environment and fatigue, and the real-time performance and consistency requirements of a high-speed production line are difficult to meet. Single modality detection (e.g., RGB only modality) is sensitive only to certain defect types, but has limited detection capability for occlusion, reflection, material difference anomalies. With the development of computer vision algorithms, the vision detection method based on deep learning can realize more accurate detection and classification through data learning and model training, and has better robustness and universality. However, supervised defect detection often relies heavily on large-scale accurate labeling data in actual landing, and abnormal sample acquisition and labeling are costly and time-consuming, and the generalization capability in the face of new types or complex defects still needs to be improved. Disclosure of Invention The invention provides a defect detection method, device, equipment and medium for medical consumables, which can expand the defect detection coverage while reducing defect labeling dependence by combining unsupervised learning with multi-mode fusion to detect the defects of the medical consumables, and is beneficial to improving the accuracy, robustness and applicability of defect detection. According to an aspect of the present invention, there is provided a defect detection method for medical consumables, the method including: acquiring multi-mode data of defect-free medical consumables as normal sample data, wherein the multi-mode data comprises visible light images, infrared images and three-dimensional point cloud information; adding interference to the normal sample data to generate pseudo-defect sample data, and determining multi-mode training data pairs according to the normal sample data and the pseudo-defect sample data; extracting features of the multi-modal training data pairs to obtain multi-modal feature information, and carrying out feature alignment on the multi-modal feature information through contrast learning to obtain target feature information; and inputting the target characteristic information into a defect detection model for self-supervision training so as to detect the defects of the medical consumable to be detected by using the trained defect detection model, wherein the defect detection model detects the defects based on the difference value between the target characteristic information and the reconstruction result. According to another aspect of the present invention, there is provided a defect detection device for medical consumables, the device including: the normal sample acquisition module is used for acquiring multi-mode data of the medical consumable without defects as normal sample data, wherein the multi-mode data comprises visible light images, infrared images and three-dimensional point cloud information; The pseudo-defect sample construction module is used for adding interference to the normal sample data to generate pseudo-defect sample data, and determining multi-mode training data pairs according to the normal sample data and the pseudo-defect sample data; the feature alignment module is used for extracting features of the multi-mode training data pairs to obtain multi-mode feature information, and carrying out feature alignment on the multi-mode feature information through contrast learning to obtain target feature information; the model training module is used for inputting the target characteristic information into a defect detection model for self-supervision training so as to detect defects of the medical consumable to be detected by using the trained defect detection model, wherein the defect detection model detects the defects based on the difference value between the target characteristic information and the reconstruction result of the target characteristic information. According to another aspect of the present invention, there is provided an electronic apparatus including: The medical consumable-oriented defect detection m