Template reconstruction method based on self-attention mechanism

Abstract

The invention provides a template reconstruction method based on a self-attention mechanism. The method comprises the steps of firstly, performing layer-by-layer down-sampling on an input image to obtain a down-sampled image; performing feature extraction on the down-sampled image; then respectively performing table look-up mapping on features of each layer by adopting a self-attention mechanism to obtain a residual error corresponding to each layer, finally, performing layer-by-layer up-sampling on the downsampled image, and meanwhile, fusing residual error information of each layer to generate an up-sampled image corresponding to each layer, thereby finally generating an up-sampled image, namely a reconstructed template image.

Description

technical field [0001] The invention belongs to the field of machine vision detection, and in particular relates to a template reconstruction method based on a self-attention mechanism. Background technique [0002] The defect detection method based on template comparison is widely used in the field of machine vision detection because of its advantages of fast speed and high precision. In the defect detection process based on template comparison, in order to improve the detection rate and reduce the false detection rate, it is often necessary to perform template reconstruction preprocessing operations, so as to eliminate the individual differences of the objects to be detected. Generally speaking, template reconstruction can be divided into rigid transformation methods and non-rigid transformation methods. Non-rigid transformation methods have received more and more attention in recent years because of their wider applicable scenarios. [0003] The non-rigid transformation ...

AI Technical Summary

Problems solved by technology

[0006] 1. In the deep self-encoding network, the low-dimensional features extracted by the encoding network generally cannot contain all the detailed features that are conducive to reconstruction, so the reconstructed image often presents varying degrees of blur, which leads to false positives in detection

Although methods such as Generative Adversarial Networks (GAN) have been proposed in recent years to alleviate the blurring problem, their principles are based on statistical information from other training images rather than the image to be reconstructed itself, so it cannot completely solve the problem of false positives

[0007] 2. For the decoding network, the operation of transposed convolution (transposed-conv / deconv) is usually used for feature upsampling. Since transposed convolution is a pathological problem, checkerboard artifacts often occur in the final reconstruction result. , thus affecting the detection results

Although methods such as subpixel convolution (subpixel conv) have been proposed to alleviate this phenomenon, they still cannot eradicate its influence.

[0008] 3. Due to the poor interpretability of the convolutional neural network itself, it is difficult to feed back from the observation of the reconstruction effect to the design of the deep autoencoder network. The selection of hyperparameters such as low-dimensional feature dimensions and the number of convolution kernels is usually Based on empirical settings, it is difficult to achieve the optimal balance in effect and efficiency, and it also hinders the transfer of reconstruction methods between different scene problems

Images