Search

CN-122023355-A - Silica gel component screening analysis system based on image recognition

CN122023355ACN 122023355 ACN122023355 ACN 122023355ACN-122023355-A

Abstract

The invention relates to a silica gel component screening analysis system based on image recognition, which belongs to the technical field of image recognition and material analysis intersection, and comprises a multi-scale microscopic image acquisition, global context perception feature extraction, weak feature enhancement and noise suppression, component semantic segmentation, quantitative analysis and confidence evaluation and self-adaptive threshold decision unit; the global context perception feature extraction module abandons the locality limitation of the traditional convolution, and utilizes an axial attention mechanism to respectively establish long-range dependence in the horizontal and vertical directions, so that the model can perceive the spatial association of weak impurities in the whole view field, and the identification capability of isolated and low-contrast particles is remarkably improved.

Inventors

  • HUANG LINGZHI
  • LI DAN
  • YANG XIUSHENG

Assignees

  • 东莞市雄驰电子有限公司

Dates

Publication Date
20260512
Application Date
20260202

Claims (10)

  1. 1. A silica gel component screening analysis system based on image recognition, comprising: The multi-scale microscopic image acquisition module is used for acquiring high-resolution microscopic image sequences of the silica gel sample under different magnification factors, registering and fusing the image sequences to generate a full-scale fused image with enhanced contrast and detail retaining capacity; The global context perception feature extraction module is used for carrying out feature coding on the full-scale fusion image, and the module adopts a non-local feature interaction structure based on an axial attention mechanism to respectively construct long-distance dependency relations along the horizontal direction and the vertical direction so as to capture the spatial distribution context information of the tiny impurities in the whole image; the weak feature enhancement and noise suppression module is used for receiving the feature map output by the global context sensing feature extraction module, dynamically adjusting the response weight of each channel and the spatial position through a channel-space dual gating mechanism, suppressing irrelevant activation of a background area, and amplifying weak signal response of a suspected impurity area; the component semantic segmentation module is used for executing pixel-level semantic segmentation based on the feature map enhanced by the weak features, accurately outlining the spatial outline of the nano-level silicon oxide impurity and outputting a corresponding binary mask map; the quantitative analysis and confidence evaluation module is used for calculating the quantity, the area occupation ratio, the average particle size and the space distribution density of the impurity particles according to the binary mask map generated by the component semantic segmentation module, and generating component confidence scores of each particle by combining the activation intensity and the context consistency index in the segmentation process; And the self-adaptive threshold decision unit is used for dynamically setting an impurity judgment threshold according to the confidence score distribution output by the quantitative analysis and confidence evaluation module, marking particles with confidence higher than the threshold as effective silicon oxide components, and eliminating noise or artifacts which are regarded as being lower than the threshold.
  2. 2. The silica gel component screening analysis system based on image recognition according to claim 1, wherein the multi-scale microscopic image acquisition module is provided with at least three objective lenses with different magnifications, wherein the magnifications are respectively 50 times, 100 times and 200 times, the module synchronously records the illumination intensity, exposure time and focal plane position parameters of each frame of image in the acquisition process, performs space alignment on the multi-scale image by using a sub-pixel level alignment algorithm based on a phase correlation method, and generates a single full-scale fusion image by using a Laplace pyramid fusion strategy, wherein high-frequency details are dominated by 200 times of images, and a low-frequency structure is dominated by 50 times of images.
  3. 3. The silica gel component screening analysis system based on image recognition according to claim 1, wherein the global context-aware feature extraction module adopts a four-stage downsampling encoder structure, and each stage of encoder is followed by an axial attention block; The axial attention block firstly carries out one-dimensional global attention calculation along the width direction of the image to generate a row-direction context feature, then carries out one-dimensional global attention calculation along the height direction to generate a column-direction context feature, and finally carries out weighted fusion on the original feature, the row context feature and the column context feature; wherein the attention weight is generated by a learnable query-key matching function in the form of a scaled dot product whose scaling factor is the inverse square root of the feature dimension.
  4. 4. The silica gel component screening analysis system based on image recognition according to claim 1, wherein the weak feature enhancement and noise suppression module comprises a channel attention sub-module and a space attention sub-module, wherein the channel attention sub-module carries out global average pooling on an input feature map along a space dimension to obtain a channel description vector, and generates channel weights through two full-connection layers and a Sigmoid activation function; The space attention sub-module performs maximum pooling and average pooling along the channel dimension after channel compression, generates a space weight map through a 7×7 convolution kernel after the obtained features are spliced, and finally outputs the element-by-element product of the input features, channel weights and space weights.
  5. 5. The system of claim 1, wherein the component semantic segmentation module uses a U-shaped codec architecture, wherein the encoding path multiplexes the four-level feature outputs of the global context-aware feature extraction module, the decoding path upsamples by transpose convolution, and jump connections are introduced at each upsampling level to fuse features of the corresponding encoding level; The method comprises the steps that a double output head is arranged at the tail end of a decoder, an impurity probability map is generated by a first output head, a boundary sharpening map is generated by a second output head, and a final segmentation result is generated by weighted fusion of the two output heads, wherein the boundary sharpening map is used for strengthening gradient response of impurity edges, and contour closure is improved.
  6. 6. The silica gel component screening analysis system based on image recognition according to claim 1, wherein the quantitative analysis and confidence evaluation module combines three indexes when calculating the confidence of single particle impurities: Firstly, outputting average activation values of all pixels in the particle region in a component semantic segmentation module in a probability map; Secondly, defining the local contrast of the particle in the original full-scale fusion image as the absolute value of the difference between the gray average value of the particle interior and the gray average value of the surrounding annular background area; thirdly, the consistency score of the particle in the multi-scale image, namely the number of times of detection in the images of 50 times, 100 times and 200 times; The three indexes are weighted and summed after normalization, and the weight coefficients are respectively 0.5, 0.3 and 0.2.
  7. 7. The silica gel component screening analysis system based on image recognition according to claim 1, wherein the adaptive threshold decision unit performs distribution modeling on confidence scores of all particles by adopting a bimodal fitting method based on a Gaussian mixture model; and judging the lower average value of the two Gaussian components obtained by fitting as a noise component, judging the higher average value as a real impurity component, setting an impurity judgment threshold value as a confidence value at the intersection point of the two components, and if the Gaussian mixture model fails to converge, backing to a fixed threshold value of 0.65 to serve as a judgment reference.
  8. 8. The silica gel component screening analysis system based on image recognition according to claim 1, wherein the system operates in a closed-loop feedback calibration mode, and automatically pushes the particle images with confidence in the interval of 0.6 to 0.7 to the manual review interface after each batch of sample analysis is completed; The rechecking result is used as newly added marking data for fine tuning decoder parameters of the component semantic segmentation module, and the fine tuning adopts an online incremental learning strategy and only updates the last two decoding layers.
  9. 9. The silica gel component screening analysis system based on image recognition according to claim 2, wherein the multi-scale microscopic image acquisition module performs low-resolution pre-scanning on a sample before image acquisition to obtain an overall reflectivity distribution map, divides the image into a plurality of brightness regions based on the distribution map, calculates optimal illumination intensity for each region independently, and realizes real-time illumination regulation and control through a digital micromirror device.
  10. 10. The system of claim 4, wherein the 7 x 7 convolution kernel is replaced by a deformable convolution kernel, and the sampling positions are dynamically generated by a lightweight offset prediction network, and the network takes an input feature map as input, outputs horizontal and vertical offsets for each spatial position, and adaptively focuses the convolution receptive field on the irregular shape of the impurity edge.

Description

Silica gel component screening analysis system based on image recognition Technical Field The invention belongs to the technical field of intersection of image recognition and material analysis, and particularly relates to a silica gel component screening analysis system based on image recognition. Background In the fields of material science and industrial quality inspection, component analysis technology based on image recognition has become an important means for improving detection efficiency and accuracy. Along with the increase of intelligent manufacturing and high-purity material demands, higher requirements are put forward on the rapid and accurate identification of trace impurities in composite materials such as silica gel. The technology generally relies on a computer vision and deep learning model, and realizes automatic screening and quantitative analysis of specific components by extracting and classifying the characteristics of microscopic images of materials. The silica gel component screening analysis based on image recognition focuses on precisely locating and identifying minute impurity components such as nano-scale silicon oxide from high resolution microscopic images. The method has the core aim of enhancing weak characteristic signals through an algorithm, and overcoming information loss caused by factors such as small scale, low contrast and the like in the imaging process, thereby realizing the detection performance of high recall rate and low omission rate. The main flow convolutional neural network is limited by a local receptive field, and is difficult to effectively capture the context association of tiny impurities in a global image, so that the characterization capability of weak characteristic components such as nanoscale silicon oxide is insufficient, and meanwhile, the visual characteristics of tiny components in a standard image are easily submerged by background noise, so that the problem of model missing detection is further aggravated. These problems are particularly prominent in application scenes such as high-purity silica gel production, which have very strict impurity control, and the reliability and practicability of an automatic quality control system are severely restricted. Disclosure of Invention The invention aims to provide a silica gel component screening analysis system based on image recognition, which is used for solving the problems of high omission ratio and low recall rate caused by limited local receptive field of a convolutional neural network and easy inundation of weak impurity characteristics by noise in the prior art, thereby realizing high-precision and high-robustness automatic recognition and quantitative analysis of trace impurity components such as nano-scale silicon oxide in a silica gel material. The technical scheme of the invention is that the method comprises the following steps: The multi-scale microscopic image acquisition module is used for acquiring high-resolution microscopic image sequences of the silica gel sample under different magnification factors, registering and fusing the image sequences to generate a full-scale fused image with enhanced contrast and detail retaining capacity; The global context perception feature extraction module is used for carrying out feature coding on the full-scale fusion image, and the module adopts a non-local feature interaction structure based on an axial attention mechanism to respectively construct long-distance dependency relations along the horizontal direction and the vertical direction so as to capture the spatial distribution context information of the tiny impurities in the whole image; the weak feature enhancement and noise suppression module is used for receiving the feature map output by the global context sensing feature extraction module, dynamically adjusting the response weight of each channel and the spatial position through a channel-space dual gating mechanism, suppressing irrelevant activation of a background area, and amplifying weak signal response of a suspected impurity area; the component semantic segmentation module is used for executing pixel-level semantic segmentation based on the feature map enhanced by the weak features, accurately outlining the spatial outline of the nano-level silicon oxide impurity and outputting a corresponding binary mask map; the quantitative analysis and confidence evaluation module is used for calculating the quantity, the area occupation ratio, the average particle size and the space distribution density of the impurity particles according to the binary mask map generated by the component semantic segmentation module, and generating component confidence scores of each particle by combining the activation intensity and the context consistency index in the segmentation process; And the self-adaptive threshold decision unit is used for dynamically setting an impurity judgment threshold according to the confidence score distribution output by the