Image Anomaly Detection Based On Positive Samples

The automated vision inspection system is required in broad industrial scenarios of quality control.It has the advantage of meeting the requirements of high-speed and high-precision detection,and it can improve the informatization,automation,and intelligence of the production process.With the increase of application scenarios,the defect data-driven detection method is faced with the problem of data deficiency and type insufficiency,and it is challenging to detect novel defects.To settle the above limitations,the anomaly detection method based on the positive sample has gotten increasing attention in the industry.This kind of method depends on positive samples as prior knowledge,making data acquirement more convenient,and can identify the novel anomalies.The detection based on positive samples is termed anomaly detection,which is usually used in big data analysis to dig outliers.In the detection process,positive sample features are extracted to construct benchmark descriptions,instances deviating from the benchmark are judged as outliers.This paper will combine machine vision technology and anomaly detection theories to explore positive sample-based detection methods in industrial scenarios.Such kind of method has essential research significance in the inspection field and has a broad prospect in application due to its advantages of low data dependence and novel anomaly sensitivity.The core tasks of image anomaly detection are image-level anomaly detection and pixel-level anomaly region segmentation.Since all the training samples are positive without labels,the primary research issues in this scenario mainly include the following two aspects :(1)The problem of feature modeling: Exploring a description extraction approach with solid robustness to ensure the distance among positives is much smaller than that between positive and negative samples in feature space.(2)The problem of anomaly measurement: Exploring optimal distance calculation approach to maximize the differentiation between positive and negative samples.This paper mainly consists of four parts,corresponding to solving the detection problems on the metal surface,texture surface,diverse object surface,and key components on scenes.The difficulty of designing feature modeling and anomaly measurement methods under the above four detection scenarios is increasing.This paper introduces different methods to settle the detection problem under different scenarios,the robustness and accuracy of the proposed methods are constantly enhanced.The major contributions in this paper are summarized as:1.Exploring the anomaly detection method for the processed surface of metal castings with irregular shapes.A reconstruction model is proposed for the first time to extract the closed contour of the region of interest to solve the problem of accurate detection region extraction.The illumination normalization method is used to ensure that the defect-free pixels in the detection area are distributed in a concentrated range,and then the anomaly detection is carried out by constructing the template of gray-value distribution characteristic.The main contribution of this method is to solve the extraction of irregular regions and the detection of abnormal pixels inside the region.Anomalies inside irregularly shaped regions are identified by designing a proprietary template for corresponding regions.The validity of this method is verified by experiments on an automatic detection platform of engine cylinder cover.2.Exploring anomaly segmentation method for texture surface.In this method,the salient contour reconstruction method extracts the structural anomaly on the textures.Meanwhile,semantic anomalies of textures are detected based on the principle of image impainting.Various data enhancement methods are proposed in this method,and artificial anomalies are introduced on positive sample data to realize self-supervision of training.Artificial anomaly is used to improve the robustness of the reconstruction model,achieving the reasonable reconstruction from abnormal to normal textures.The contribution of this method lies in leveraging data augmentation to construct the image generation model,and the outline-to-region mapping module is innovated to extract structural anomalies.The region prediction branch is innovatively introduced into the generation model to filter the pixels in the abnormal region and realize the reconstruction task of the abnormal region.This method ensures significant reconstruction differences in anomalous regions and retains texture information of defect-free regions completely,thus improving detection robustness.3.An anomaly segmentation method is proposed based on multi-level image reconstruction and adaptive attention level transition.This method mainly solves the problem that the model cannot adjust the performance under different detection objects,and further improves the robustness of the reconstruction-based anomaly detection method.The core innovation lies in constructing a multi-level image reconstruction framework,and the adaptive detection performance optimization for different objects is achieved by leveraging the reconstruction performance differences of detail texture and global structure among different generators.In this method,a perceptual reconstruction error measurement is proposed to improve anomaly region segmentation accuracy significantly.A control parameter is used to adjust the weight of reconstruction level synchronously and error measurement scale,to realize the reconstruction performance adjustment and optimize the performance of multi-scale anomaly detection.Finally,the reconstruction uncertainty between different reconstruction levels is leveraged to adapt the control parameters.This method achieves the best performance on the MVTec AD dataset among existing anomaly segmentation methods.4.An approach of target component anomaly detection in scenes is proposed based on keypoint prediction and local feature reconstruction.In this detection scene,the characteristics of positive samples vary greatly,including camera shooting position changes,illumination changes,dirt interference,and outside structure interference.Therefore,the primary work of anomaly detection for target components is to identify detection areas.The method simultaneously achieves target region instance segmentation,target point coordinate prediction,and abnormal target region segmentation tasks by constructing a compact model set through tagged training on positive samples.Its main innovation lies in using the keypoint prediction method to solve the problem that the segmentation model can not completely recognize abnormal regions.The local feature reconstruction module was introduced to reconstruct the features describing the core semantics inside the target regions,and multi-scale channel prediction branches obtained the complete target region segmentation results.Therefore,the abnormal regions can be determined by comparing the initial segmentation results with reconstructed results.This method is highly robust to environmental uncertainties because it uses the core features of case segmentation to detect anomalies and dramatically improves detection accuracy.Meanwhile,the compact structure design ensures high computational efficiency.