Aramid paper honeycomb gluing defect detection method

[0020] The specific embodiments of the present invention will be discussed in detail below in conjunction with the drawings.

[0021] Combine figure 2 , image 3 As shown in the decoder and encoder, the present invention provides a method for detecting defects in aramid paper honeycomb gluing, including aramid paper, on which there are defects generated in the gluing process; and an industrial camera for Obtain the defect image information of aramid paper; the depth feature encoder is used to fully encode different types of different forms of defect image information; the feature decoder is used to accurately locate the encoded defect image information; the detection steps are as follows figure 1 The workflow of the detection program,

[0022] The first step is to perform different image enhancement processing on the defect image information obtained by the industrial camera to obtain different defect image samples;

[0023] In the second step, the depth feature encoder performs feature encoding on the defective image information sample described in the first step, for detecting different defects in different image defect samples;

[0024] In the third step, the depth feature encoder and the feature decoder are connected to perform pixel-based positioning of different defects in a large number of image defect samples, thereby completing the defect detection process.

[0025] First, the present invention automatically generates a large number of samples through the image enhancement algorithm, thereby reducing the dependence of machine learning on the number of samples, and reducing the pressure of preliminary preparations in actual industrial production; secondly, it solves the defects of different sizes and different aspect ratios. At the same time, the detection problem greatly improves the detection efficiency; finally, the pixel-level positioning is realized, the defect detection and positioning are accurate, and the real-time performance is high, which is faster than the mainstream manual detection in the factory.

[0026] The present invention adopts a deep learning method, increases the number of samples by enhancing the algorithm of defective image information; through the design of deep feature extraction network, fully encodes different types and different forms of defective image information, and avoids the encoding of noise and image interference; The connection design of the "feature receptive field" feature fully adapts to the detection requirements of defects of different scales and different aspect ratios; the coded defects are accurately located through the decoder, and the defect location is directly output.

[0027] 1. A solution to a large number of samples that are difficult to obtain

[0028] Enhance existing images through a variety of image processing methods. The image enhancement methods used in the present invention include: random light and dark changes, random contrast changes, random scaling, random rotation, random noise, random cropping, random translation, etc. Through image enhancement, samples that are tens of times the original image can be obtained.

[0029] 2. Solutions to the wide variety of defects, diverse shapes, large scale changes, and drastic changes in aspect ratio

[0030] In order to realize effective defect detection, the present invention designs a depth feature encoder and connects the receptive fields of different scales to detect defects of different scales and aspect ratios. Take a picture with an input size of 572*572 as an example. After being processed by the encoder, the size and quantity of the feature layer change as follows figure 2 As shown, the depth feature encoder consists of 4 blocks, each block uses 3 effective convolutions and 1 maximum pooling layer downsampling. After each downsampling, the number of feature layers is multiplied by 2, so The number of feature layers shown in the figure varies. At the same time, in order to achieve high real-time online detection, it is necessary to minimize parameter calculations and simplify the number of deep network layers, so a residual structure is added between blocks.

[0031] 3. Solutions to the problem of precise positioning of defects

[0032] In order to solve the problem of precise positioning of defects, the present invention designs a feature decoder, which is also composed of 4 blocks. Before each block starts, the size of the feature layer is multiplied by 2 through deconvolution, and the number is halved to input Take a picture with a size of 572*572 as an example, and the final feature layer size is 388*388. The feature decoder is connected with the encoder to achieve precise positioning of the corresponding scale defects. Its brief structure is as image 3 Shown.