PCB defect data generation method based on deep learning

Abstract

The invention discloses a PCB defect data generation method based on deep learning. In a generator network and a discriminator network, a BatchNorm layer, a Relu layer and a LeakyRelu layer are added,the Relu layer is used for the generator network, and the LeakyRelu layer is used for the discriminator network: an encoder is designed and is composed of three convolution layers, and the encoder extracts features from an input image; a converter is designed, wherein the converter is composed of six residual blocks and converts the feature vector from a source domain to a target domain; a decoder is designed, wherein the decoder is composed of three deconvolution layers; designing a discriminator, wherein the discriminator is composed of five convolution layers; a loss function is designed,wherein the loss function comprises four parts; a training set for model training is prepared, and a network is trained to automatically repair defects; a weight file under the optimal iteration frequency is selected for testing; the restored image and the test image are compared by using the LBP, so that the position of the defect can be found more accurately; and various defect data sets are obtained by utilizing the generated samples. The method is high in robustness, wide in application range and excellent in generation effect.

Description

technical field [0001] The invention relates to the field of PCB defect detection, in particular to a method for generating PCB defect data based on deep learning. Background technique [0002] Traditionally, surface defects have been visually detected by humans, which is subjective, expensive, inefficient and inaccurate. At present, the detection and classification of surface defects based on machine vision can significantly improve the efficiency of industrial production. Machine vision systems are superior to human vision, but there are still many problems and challenges in practical production applications. Since traditional image features used to distinguish defects from non-defects are manually designed based on experience, the features of traditional image feature extraction algorithms are usually at a low level, under complex production scene changes, such as illumination changes, perspective distortion, There are partial occlusions, object deformation, etc., and t...

AI Technical Summary

Problems solved by technology

Downsampling of non-defective samples, during batch training, limit the number of non-defective samples in each batch of randomly drawn samples

In the PCB defect detection project, the number of open-circuit samples with the least amount of data is only thousands, while the number of samples without defects is on the order of 100,000. The number of category samples is seriously skewed. lead to overfitting of the model

Images