A cell defect classification method based on unbiased embedding zero-sample learning

Abstract

A method for classify cell defects base on unbiased embedded zero-sample learning includes such steps as 1) obtaining an image block containing defects with suitable size from a cell picture collectedby a high-definition camera through a sliding window, allocating common defect categories as labeled source data, and allocating rare defect categories as unlabeled target data; 2) end-to-end training of the quasi-fully supervised learning network model QFSL using annotated source class data and unannotated target class data; (3) In the testing phase, the input image block is embedded by the visual embedding subnetwork, and then the image block is embedded by the visual embedding subnetwork, and the image block is embedded by the visual embedding subnetwork. Semantic linking sub-network completes the mapping from visual embedding to semantic embedding, then gets the scores of visual embedding and semantic embedding by inner product calculation, and finally sends them to Softmax classifierto generate the prediction probability of all categories, and takes the category with the highest probability as the classification result.

Description

technical field [0001] The invention belongs to the field of industrial defect identification. Aiming at the problem that the data of some defects in the battery defect classification scene is relatively small or difficult to obtain, a battery defect classification method based on unbiased embedded zero-sample learning is proposed. Background technique [0002] Cell defect identification is an important part of the battery production process. In the modern society with the rapid development of industrialization, the traditional way of using manual inspection has many disadvantages: (1) It is labor-intensive. The method of manually checking for defects requires a lot of manpower and time to match the needs of the battery market; (2) The evaluation lacks objectivity. The manual judgment of defects mainly depends on personal experience, which is difficult to unify for different workers, so the final judgment is relatively subjective; (3) the visual impairment of workers is gre...

AI Technical Summary

Problems solved by technology

In the battery defect scenario, the main defect categories include pits, single-layer separator damage, separator folds, scratches, tab discounts, glue stains, damaged metal leakage, air bubbles, toner sticks, and toner dirt. There are 13 categories of dirt, crushing, foreign matter, and oil pollution, some of which have a relatively low probability of occurrence, and the corresponding sample size is too small, which is not enough to be directly used for the training of the defect classifier

In addition, some defects are gradually excavated with the classification, and new types of defects will appear

Classification in this case requires a zero-shot learning strategy, but most of the existing zero-shot learning methods suffer from strong bias: during the training phase, the input is usually projected to several parts of the semantic embedding space determined by the source class. fixed point

Images