Current location - Education and Training Encyclopedia - Graduation thesis - Clipping Interpretation: Self-supervised Learning for Anonymous Detection and Location
Clipping Interpretation: Self-supervised Learning for Anonymous Detection and Location
This paper aims to solve the problem of single classification. Through the study of normal samples, unknown types of anomalies are detected, and defect detection is more inclined. Methods The whole process was divided into two steps. Firstly, the sample representation is learned based on the self-supervised learning method, and then a single classifier is used to classify the sample representation. Through the CutPaste proposed in this paper, the representation learning and classification of normal samples are realized CutPaste is a data enhancement method, which cuts a rectangular area from an image and pastes it at any position. The purpose of cutting and pasting is to produce spatial irregularity as an approximation of the real defects that do not participate in training. Some popular existing methods apply rotation and contrast learning to a single classification, but the experiments in this paper prove that rotation or contrast learning is not optimal for defect detection. The author suspects that geometric transformations, such as rotation and displacement, are effective in learning the representation of semantic concepts (such as learning objects), but

Learn less about regularity (for example, continuity and repetition). For defect detection, this paper hopes to propose a data enhancement method to simulate local irregular patterns.

At present, the main idea of single classification anomaly detection algorithm is to train a model that can represent normal samples and assume that this model can not represent abnormal samples well. However, high-level semantic information cannot be obtained based on pixel-level reconstruction loss.

This paper adopts the idea of subterfuge task in self-supervised learning. In the characterization stage of self-supervised learning, cutpaste is used to generate images for positive samples, and a binary CNN is trained to identify normal samples and images added with cutpaste. In the stage of (b) anomaly detection and location, CNN is used to extract features, and parameter Gaussian probability density estimation (GDE) uses the features extracted by CNN to calculate anomaly scores. For image-level anomaly detection, we can use GradCAM to roughly locate the anomaly area, and use patch-level anomaly location to divide the original image into several patches and send them to CNN-GDE to calculate the anomaly scores respectively, so as to get a finer-grained anomaly heat map.

(The second stage should be the abnormal score calculated directly through the representation of the output without training. )

Found a phenomenon.

Those who use contrastive self-supervised learning for anomaly detection (simclr et al. The second stage basically needs to do finetune.

If other self-supervised learning is used for anomaly detection (which means that the subterfuge task does not use contrast learning, but identifies whether the image is rotating), fine tuning may not be performed in the second stage.

The data enhancement methods used in this paper, cutpaste and cutpaste(scar), are inspired by cutout and scar, and scar is to add slender lines. The actual self-supervised learning experiment is regarded as three classification problems: normal samples, samples with cutpaste and samples with cutpaste(scar).

For MVTec AD data sets, or actual defect detection tasks, defects generally include tensile deformation and special texture. The purpose of using cutpaste for normal samples is to simulate abnormal samples by using cutpaste. Through the use of visualization technology, the author found that the distance between the positive sample and the original positive sample increased, but the proximity to the real abnormal sample decreased, which indicated that better data enhancement methods were still needed.

Experimental comparison

Experimental comparison, defect location

Ablation experiment with different data enhancement methods

Learn and evaluate the representation of depth-classification. In the self-supervised learning stage, different angles of image rotation are regarded as different categories. )

The author thinks that for different types of data sets, subterfuge tasks combined with different data enhancement methods have great influence. For semantic data sets, the rotation effect is better, and for detail defect detection, the performance of this paper is better. It is necessary to design data enhancement methods according to the characteristics of anomalies, so as to enrich samples which are quite different from positive samples, and it is best to display the characteristics of abnormal samples.