CN-121999486-A - Diatom image automatic identification and water quality evaluation system

CN121999486ACN 121999486 ACN121999486 ACN 121999486ACN-121999486-A

Abstract

The invention belongs to the technical field of automatic image recognition and processing, and discloses an automatic diatom image recognition and water quality evaluation system which comprises an image normalization engine, a recognition engine and a residual error quantification module, wherein the image normalization engine is used for training based on paired training data sets manually cleaned and repaired by an expert and is used for converting an original microscope view field image containing semantic interference into a normalized cleaning image, the recognition engine is connected with the image normalization engine in series and is used for only receiving the normalized cleaning image for classification and counting, and the residual error quantification module is connected with the image normalization engine in parallel and is used for receiving the residual error map and carrying out quantification statistics so as to evaluate physical indexes of water quality. According to the invention, through a serial architecture, the semantic cleaning task and the standard recognition task are decoupled, so that the standard recognition engine operates under the input condition close to ideal, and the technical problem that the standard model is invalid in recognition due to an interfering object in an actual field environment is solved.

Inventors

YANG LIN
WANG YUE
LIU HAIPENG
WU SHIKAI

Assignees

衡水学院

Dates

Publication Date: 20260508
Application Date: 20251209

Claims (9)

1. The diatom image automatic identification and water quality evaluation system is characterized by comprising an image normalization engine, an identification engine and a residual error quantization module; The image normalization engine is used for providing a technical premise of semantic normalization for the operation of the recognition engine, the technical premise is realized by the following mechanism that the image normalization engine is used for receiving an original microscope view field image, the original microscope view field image comprises target diatom and semantically non-target interferents, the non-target interferents comprise non-diatom fragments, fuzzy contours outside a focal plane and broken diatom shells, the image normalization engine is used for training a pair of training data sets, the pair of training data sets comprises the original microscope view field image serving as an input and an object image which is manually cleaned and repaired by an expert corresponding to the original microscope view field image and is output as a target, the image normalization engine is used for converting the original microscope view field image into a normalized cleaning image which inhibits the characteristics of the semantically non-target interferents based on the training operation, and the image normalization engine is further used for generating a normalized residual map based on the difference of image characteristics between the original microscope view field image and the normalized cleaning image; The operation of the recognition engine is constrained by the premise of semantic normalization technology, and the constraint is embodied by a mechanism that the recognition engine is connected with the output end of the image normalization engine in series and is used for only receiving normalized cleaning images so as to classify and count target diatoms in the cleaning images and generate a diatom recognition result containing target diatom categories and corresponding confidence scores; The residual quantization module is connected in parallel to the output end of the image normalization engine and is used for receiving the normalized residual spectrum, carrying out quantization statistics on the normalized residual spectrum, and generating a quantization index representing the total amount of non-target interference substances and used for realizing water quality evaluation.
2. The diatom image automatic identification and water quality evaluation system according to claim 1, wherein the image normalization engine is operated based on an image-to-image translation model, the image-to-image translation model adopts a full convolution neural network architecture, and the full convolution neural network architecture is used for learning semantic mapping rules for training the data set in pairs, from an original microscope field-of-view image to a target image after manual cleaning by an expert.
3. The automated diatom image identification and water quality assessment system of claim 1, wherein the image normalization engine is operated based on an image-to-image translation model employing a generating countermeasure network architecture for generating normalized cleaning images that are semantically consistent in image characteristics with the human expert manually cleaned and restored target images.
4. The system for automatically identifying diatom images and evaluating water quality according to claim 1, further comprising an output monitoring module connected to an output end of the identification engine and used for monitoring confidence scores in diatom identification results in real time, a tracing archiving module linked with the output monitoring module and having data access rights to the original microscope field images, wherein the tracing archiving module is triggered when the output monitoring module detects that the confidence scores are lower than a preset judgment threshold value and used for automatically grabbing and archiving the original microscope field images corresponding to the low confidence scores.
5. The automated diatom image identification and water quality assessment system of claim 1, wherein the image normalization engine is configured to operate in a random inference mode by activating its internal Dropout layer to generate N sets of normalized cleaning images for the same raw microscope field of view image, N being an integer greater than 1, the identification engine is configured to traverse the N sets of normalized cleaning images to generate N sets of corresponding diatom identification results, and the system further comprises a consensus aggregation module configured to receive the N sets of diatom identification results and to determine a single final diatom identification result based on a preset consensus logic.
6. The system for automatically identifying diatom images and evaluating water quality according to claim 5, wherein the preset consensus logic is a confidence weighted average logic, and the consensus aggregation module is configured to obtain, for any target diatom category in the N sets of diatom identification results, a confidence score of the category in each set of diatom identification results, calculate an average value of the N sets of confidence scores to obtain a final consensus confidence, and output a final diatom identification result based on the final consensus confidence.
7. The system for automatically identifying diatom images and evaluating water quality according to claim 1, further comprising an output monitoring module connected to an output end of the identification engine for monitoring the confidence score in the diatom identification result in real time, and a bootstrap filing module, in linkage with the output monitoring module, wherein the bootstrap filing module is triggered when the output monitoring module detects that the confidence score is higher than a preset high confidence threshold, and is used for automatically capturing and filing the original microscope field image corresponding to the high confidence score and the diatom identification result generated by the identification engine as a data pair for model bootstrap training.
8. The automated diatom image identification and water quality assessment system of claim 1, further comprising a process audit module for visual output of the superposition of the normalized residual spectrum and the diatom identification results generated by the identification engine.
9. The automated diatom image recognition and water quality assessment system of claim 1, further comprising a color reference storage module for storing a reference color spectrum corresponding to a standard color space employed during training of the image normalization engine, and a color normalization gateway in series with an input of the image normalization engine, wherein the color normalization gateway is configured to perform a color histogram matching algorithm to match the color space of the original microscope field image with the reference color spectrum prior to processing of the original microscope field image by the image normalization engine.

Description

Diatom image automatic identification and water quality evaluation system Technical Field The invention relates to an automatic diatom image recognition and water quality evaluation system, and belongs to the technical field of automatic image recognition and processing. Background Current deep learning based image recognition models, such as convolutional neural networks, both training and running rely on clear, standardized input images with high accuracy in processing such standardized data sets, however, when these trained image recognition models are deployed into field reality conditions in laboratory environments, their performance is degraded, mainly because the raw images acquired in the field are often mixed with large amounts of non-target complex interference information, which are not traditionally pixel noise, but have semantically-level interference features, such as non-target organic debris, inorganic particles, bubbles, broken shells that are prevalent in raw objective lens inspection, in water samples, and blurred or semitransparent contours at different focal planes due to depth of field limitations. To solve this problem, one existing idea is to try to build a more powerful single recognition model, let the model learn all possible interference patterns in training, try to learn to directly recognize the target from the image containing the interference; however, in other related water quality detection fields, there are also technical schemes attempting to avoid image interference by extracting highly simplified statistical features, but there are fundamental limitations, for example, chinese patent application publication No. CN114264648B discloses a water quality detection test paper identification method and system, the scheme aims to identify relatively uniform color development areas, the technical idea is to use statistical feature data such as mean value and standard deviation of R, G, B single channel as input of neural network, the processing mode fundamentally discards image morphology and spatial distribution information, however, in the microscopic examination scene of the invention, the target object and non-target object (such as debris and non-target object) have severe morphological feature-based on the statistical feature identification, the invention aims to identify the complex morphology and the complex morphology of the diatom, the method is not applicable to the technical problem of target identification under complex semantic interference to be solved by the method. Therefore, how to design an image recognition system architecture, the architecture can logically separate a complex semantic interference processing task from a standard target recognition task, so that a standard recognition model can run on an image with interference suppressed, and the problem of recognition performance reduction caused by on-site image interference is solved. Disclosure of Invention The invention provides an automatic diatom image recognition and water quality evaluation system, and mainly aims to provide an image recognition system architecture which can logically separate complex semantic interference processing tasks from standard target recognition tasks so as to solve the problem of recognition performance reduction caused by field image interference. In order to achieve the above purpose, the invention provides a diatom image automatic identification and water quality evaluation system, which comprises an image normalization engine, an identification engine and a residual error quantization module; The image normalization engine is used for providing a technical premise of semantic normalization for the operation of the recognition engine, the technical premise is realized by the following mechanism that the image normalization engine is used for receiving an original microscope view field image, the original microscope view field image comprises target diatom and semantically non-target interferents, the non-target interferents comprise non-diatom fragments, fuzzy contours outside a focal plane and broken diatom shells, the image normalization engine is used for training a pair of training data sets, the pair of training data sets comprises the original microscope view field image serving as an input and an object image which is manually cleaned and repaired by an expert corresponding to the original microscope view field image and is output as a target, the image normalization engine is used for converting the original microscope view field image into a normalized cleaning image which inhibits the characteristics of the semantically non-target interferents based on the training operation, and the image normalization engine is further used for generating a normalized residual map based on the difference of image characteristics between the original microscope view field image and the normalized cleaning image; The operation of the recognition engine is constrained by the p