CNN (Convolutional Neural Network) visualization-based PCB (Printed Circuit Board) defect detection method

Abstract

The invention discloses a CNN (Convolutional Neural Network) visualization-based PCB (Printed Circuit Board) defect detection method, which overcomes the problem of poor interpretability of a convolutional neural network, corrects training data errors by utilizing a CNN model and combining model visualization of the CNN model, and checks the accuracy of model category positioning to improve the precision of defect detection. The method comprises the following steps: training acquired PCB image data through a convolutional neural network; then positioning a thermodynamic diagram through Grad-CAM categories to judge whether the basis for predicting the visualization model is correct or not, detecting the cause of a data error of a prediction error, and then adjusting the data to reduce the error of a training set, thereby achieving the purpose of improving the precision of the data set.

Description

technical field [0001] The invention relates to the technical field of deep learning computer vision, in particular to a method for detecting defects of PCB circuit boards based on CNN visualization. Background technique [0002] For a long time, CNN has been controversial despite its remarkable effect. The root cause is its poor interpretability. How did the neural network produce this result? Especially when the number of layers of the neural network is large, the interpretability is very poor, and the working principle of the deep neural network cannot be described in words. Therefore, before the emergence of convolution visualization, it has been regarded as a "black box". But the reality is that our users or terminals need interpretability, so the research field of convolution visualization has been derived. [0003] The Grad-CAM (Gradient-weighted Class Activation Mapping) method can make the CNN model interpretable and transparent. By calculating the neuron importanc...

AI Technical Summary

Problems solved by technology

[0002] For a long time, CNN has been controversial despite its remarkable effect. The root cause is its poor interpretability. How did the neural network produce this result? Especially when the number of layers of the neural network is large, the interpretability is very poor, and the working principle of the deep neural network cannot be described in words. Therefore, before the emergence of convolution visualization, it has been regarded as a "black box".

Images