Liquid crystal display panel CF picture identification method based on statistical learning

Abstract

The invention discloses a liquid crystal display panel CF picture identification method based on statistical learning, and relates to the field of image target detection. The liquid crystal display panel CF picture identification method includes the steps: abstracting a question for identifying defect of a liquid crystal panel as a target detection question and a classification question in image processing; performing exploratory data analysis; annotating images, and deleting the images which are not beneficial to training to avoid the interference on a model; preprocessing the data; performing data enhancement on the images: performing data enhancement by transforming an original image, wherein the transformation mode comprises the steps of randomly overturning the image, changing the color of the image and the like; training a convolutional neural network model, improving the current specific question, and carrying out model evaluation; and outputting a final model, and detecting andclassifying the liquid crystal panel. For the liquid crystal display panel CF picture identification method, pictures are out of order before pre-reading, so as to avoid excessive memory space consumption caused by out-of-order during pre-reading while preventing only one type of data from being used for weight updating each time, and improving the stability of the model training process; and thelearning rate is dynamically adjusted according to image characteristics, so that the convergence process is accelerated, and the number of iterations is reduced.

Description

technical field [0001] The invention relates to the field of image target detection, in particular to a method for recognizing CF pictures of liquid crystal panels based on statistical learning. Background technique [0002] The types of CF pictures in the process of generating LCD panels are mainly divided into ITO (including FI312, FI609, FI611) and BM (including FM307, FM312, FM601). According to the industry's production process, the ITO and BM information is known before inspection, so the two types of pictures are subjected to target detection. The purpose of target detection is to classify damaged pictures and frame the damaged location. [0003] The current mainstream target detection methods are mainly divided into two categories: [0004] (1) Two-stage method: RCNN architecture [0005] In 2014, it was originally proposed by Ross Girshick et al. at the Facebook AI Research Institute. [0006] Features: High precision, slow speed, mostly used for academic researc...

AI Technical Summary

Problems solved by technology

[0035] The purpose of the present invention is to provide an improved target detection algorithm based on YOLOv3 for the classification and target detection of errors and defects generated in the CF process in the production of liquid crystal panels, which can improve the overall classification accuracy of the model and speed up training and convergence. At the same time, it solves many problems such as small samples and large training fluctuations, so that the classification success rate of most categories is increased to more than 85%, a CF image recognition method based on statistical learning for LCD panels

Images