Microplastic multi-dimensional characteristic high-throughput rapid detection method

Abstract

The invention discloses a microplastic multi-dimensional feature high-throughput rapid detection method, which is characterized in that microplastic multi-dimensional feature synchronous detection is carried out by coupling microimaging, Raman spectrum, deep learning and machine learning, so that full-process automatic processing, multi-modal data fusion synchronous characterization, standardized data integration and report generation are realized; and a structured report containing morphological characteristic parameters can be automatically output, cross-research data comparison is supported, and detection standardization is promoted. Meanwhile, the detection efficiency is remarkably improved; the detection time of a single sample is shortened to 15 minutes; the detection precision is comprehensively guaranteed: the material identification accuracy is greater than or equal to 95%; the form classification accuracy is greater than or equal to 92%; a volume prediction average error lt; 40%; and multi-dimensional synchronous analysis capability: synchronously outputting a plurality of core morphological characteristics in a single process for the first time, and avoiding sample loss and data deviation caused by step-by-step detection. And a reliable tool is provided for micro-plastic pollution mechanism research and environment monitoring, so that the method has a wide application prospect.

Description

Technical Field

[0001] This invention belongs to the field of microplastic detection technology, and specifically relates to a high-throughput rapid detection method for multidimensional features of microplastics. Background Technology

[0002] Microplastics (plastic fragments with a particle size <5 mm) are emerging environmental pollutants that have been widely detected in aquatic, terrestrial, and biological systems worldwide, and their ecological risks and health threats urgently require accurate assessment. Traditional microplastic detection methods rely heavily on manual intervention, and the labor-intensive operation results in insufficient efficiency and accuracy. For example, in the morphological identification stage, manual microscopic observation easily misidentifies clay, organic debris, etc., as microplastics, with a misclassification rate of over 30%, and it is only applicable to particles with a particle size >1 mm; Raman spectroscopy or FTIR analysis requires scanning particles point by point, and the detection of a single filter membrane takes more than 2 hours, which is difficult to meet the needs of high-throughput analysis of environmental samples and seriously restricts analytical efficiency and data reliability.

[0003] Acquiring multidimensional features relies on combinations of multiple instruments, making simultaneous characterization difficult. Key characteristics of microplastics encompass morphology (particle size, shape, quantity), chemical composition (material), and physical properties (volume), requiring step-by-step detection with current technologies. For morphology and composition detection, the mainstream method involves first obtaining morphological parameters using a microscope, then transferring the sample to a spectrometer for compositional analysis—a cumbersome process prone to particle loss or contamination. Furthermore, volume is a core parameter for assessing the environmental migration and bioavailability of microplastics, but current methods can only estimate based on the spherical assumption, resulting in errors exceeding 30% for irregular particles (such as fibers and fragments). Direct measurement methods (such as layered imaging systems) also have a time cost exceeding 5 minutes per sample for a single particle, making them inefficient.

[0004] In addition, existing technologies suffer from low levels of data integration and standardization. Data generated by multiple instruments requires manual integration and cannot simultaneously output related features such as morphology, material, and volume, making it difficult to quickly form a multidimensional feature database of microplastics to support subsequent analysis. Summary of the Invention

[0005] To address the problems existing in the prior art, the present invention aims to provide a high-throughput, rapid detection method for multidimensional features of microplastics. This method for simultaneous detection of multidimensional features automates the process, replaces manual operation, achieves simultaneous characterization through multimodal data fusion, and standardizes data integration and report generation. It automatically outputs structured reports containing parameters such as particle ID, particle size, shape, material, and volume, supports cross-study data comparison, and promotes detection standardization. This method overcomes the efficiency, accuracy, and dimensionality limitations of traditional detection methods, providing a reliable tool for research on microplastic pollution mechanisms and environmental monitoring.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: This invention provides a high-throughput rapid detection method for multidimensional features of microplastics, comprising the following steps: S1. Obtain a microscopic image and corresponding coordinates of the microplastic sample, and obtain the Raman spectral data at that coordinate location; then, based on the obtained Raman spectral data, determine the material of the microplastic through spectral matching. S2. Perform automated image processing on the microscopic images obtained in step S1 to identify microplastic particles and extract morphological features, including particle size, shape, shape factor and number of particles. Preferably, step S1 specifically includes the following steps: S11. Use a low-power microscope to obtain microscopic images of the sample, and identify microplastic particles by grayscale threshold segmentation to obtain the coordinate positions of the microplastic particles. S12. Automatically position the Raman spectrometer to the coordinate position, collect Raman spectral data, and determine the microplastic material based on the Raman spectral data through spectral matching. S13. Perform spectral matching between the obtained Raman spectral data and the preset spectral library based on the characteristic peak position and intensity similarity threshold to determine the material of the microplastic.

[0007] Preferably, step S2 specifically includes the following steps: inputting the microscopic image obtained in step S1 into a pre-trained DeepLabV3 neural network to perform precise contour segmentation of the microplastic particles in the microscopic image and output a binary mask; then filtering out impurity interference through morphological operations, using OpenCV to traverse the mask contour, identifying the microplastic particles and calculating and extracting the morphological feature parameters of each particle, including the equivalent circle diameter of each particle as the particle size, the aspect ratio AR as the shape parameter, the roundness C as the shape factor, and the number of particles, and automatically classifying according to the aspect ratio AR and roundness C thresholds to determine the shape of the microplastic particles.

[0008] Preferably, the principle for automatically classifying and determining the shape of microplastic particles based on the aspect ratio AR and the roundness C threshold is as follows: when AR>5, the shape of the microplastic particles is fibrous; when C<0.7 and AR≤5, the shape of the microplastic particles is fragmented; when C<0.7 and AR≤5, the shape of the microplastic particles is spherical.

[0009] S3. Construct a volume prediction model based on PyCaret: Based on the morphological feature parameters of each particle, establish a machine learning model, wherein the model takes shape factor, particle size and particle area as input features, takes the measured volume as the training target, and outputs the predicted microplastic volume. Preferably, the machine learning model is a volume prediction model constructed using a random forest regression algorithm, and the training sample set includes measured volume data of microplastics of different materials and corresponding morphological feature parameters such as shape factor, particle size, and particle area.

[0010] Preferably, the machine learning model is trained in the following way: collecting microplastic samples of known volume, obtaining their microscopic images and Raman data, extracting shape factor, particle size, and particle area features, using the measured volume as a label, traversing the machine learning regression models in the Pycaret library, and selecting the optimal model for predicting the volume of microplastic particles.

[0011] S4. Multidimensional feature integration output of microplastics, including microplastic particle ID, particle size, shape category, area A, aspect ratio AR, solidity S, roundness C, predicted volume V, and material.

[0012] Preferably, it also includes: microplastic abundance (particle count), proportion distribution of each material, volume distribution, and shape proportion distribution.

[0013] Preferably, in step S12, the excitation wavelength of the Raman spectrometer is 785 nm and the integration time is 1-5 s.

[0014] Beneficial effects This invention proposes a high-throughput, rapid detection method for multidimensional features of microplastics. It is a synchronous detection method for multidimensional features of microplastics based on a coupling of microscopic imaging, Raman spectroscopy, deep learning, and machine learning. This method automates the entire process, replacing manual operation, and simultaneously characterizes multimodal data through fusion, standardizing data integration and report generation. It can also automatically output structured reports containing parameters such as particle ID, particle size, shape, material, and volume, supporting cross-study data comparison and promoting detection standardization. Furthermore, it achieves the following beneficial effects: (1) Significantly improved detection efficiency: The detection time for a single sample is reduced to 15 minutes (more than 5 times faster than traditional methods). (2) Comprehensive guarantee of detection accuracy: material identification accuracy ≥95% (Raman spectroscopy matching threshold optimization); morphology classification accuracy ≥92% (DeepLabV3 segmentation + shape factor constraint); average volume prediction error <40% (ensemble regression model); (3) Multidimensional synchronous analysis capability: For the first time, it realizes the synchronous output of core morphological features such as particle size, shape, material and volume in a single process, avoiding sample loss and data deviation caused by step detection.

[0015] In summary, this invention breaks through the limitations of traditional detection methods in terms of efficiency, accuracy, and dimensionality through technological innovation, providing a reliable tool for the study of microplastic pollution mechanisms and environmental monitoring, and therefore has broad application prospects. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 A flowchart illustrating a high-throughput rapid detection method for multidimensional features of microplastics provided by this invention; Figure 2 A representative image showing the detection profile of microplastic particles; Figure 3 Results of environmental microplastic detection and feature extraction. Detailed Implementation

[0018] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of the invention. However, those skilled in the art will understand that the invention may be implemented in other embodiments without these specific details.

[0019] Example 1 Reference Figure 1 A high-throughput rapid detection method for multidimensional features of microplastics includes the following steps: S1. Obtain a microscopic image of the microplastic sample and its corresponding coordinates, and obtain the Raman spectral data of the coordinate point; based on the obtained Raman spectral data, determine the material of the microplastic through spectral matching; Specifically, large-field microscopic imaging and Raman spectroscopy are performed in tandem, including: (1) A microscopic panoramic image of the sample area was obtained using a low-power microscope (4× objective or 10× objective), and suspicious microplastic particles were identified by gray-scale threshold segmentation (threshold = mean + 1.5× standard deviation, here threshold = 130, range 0-255) to obtain the coordinate position of the microplastic particles; (2) Automatically position the Raman spectrometer (excitation wavelength 785 nm, integration time 1-5 s) to the coordinate position and collect Raman spectral data; Based on the collected Raman spectral data, the material of the microplastic is determined by spectral matching. Specifically, determining the material of the microplastic by spectral matching includes: performing spectral matching between the collected Raman spectral data and a preset spectral library (containing characteristic spectra of 30 common microplastics) based on the similarity threshold (≥80%) of characteristic peak positions and intensity to determine the material of the microplastic. S2. Perform automated image processing on the microscopic images obtained in step S1 to identify microplastic particles and extract morphological features, including particle size, shape, shape factor and number of particles. Specifically, the microscopic images obtained in step S1 undergo automated image processing, specifically particle contour recognition and morphological analysis based on DeepLabV3. This includes: inputting the acquired microscopic images into a DeepLabV3 neural network (backbone network ResNet-50, output stride = 16), training the model using a weighted sum of cross-entropy loss and Dice coefficient loss (weights 0.5:0.5), performing precise contour segmentation of the microplastic particles, and outputting a binary mask of the microplastic particles; then filtering out impurity interference through morphological operations, using OpenCV to traverse the mask contours, identifying the microplastic particles, and calculating and extracting the morphological feature parameters of each particle, specifically including: Equivalent circle diameter: as particle size; Particle area (A): Number of outline pixels × Calibration factor (μm) 2 / pixel); Aspect Ratio (AR): The ratio of the major axis to the minor axis of the smallest bounding rectangle, and used as a shape parameter; Solidity (S): Contour area / Convex hull area; Circularity (C): 4πA / circumference 2 And use it as a shape factor; and the number of particles; Then, based on the aspect ratio (AR) and roundness (C) thresholds, the shape of the microplastic particles is automatically classified and determined: fibrous (AR>5), fragmented (C<0.7 and AR≤5), and spherical (C≥0.7 and AR≤5).

[0020] S3. Based on the morphological feature parameters of each particle, a machine learning model is established, wherein the model takes shape factor, shape parameter and particle area as input features, and the measured volume as the training target. The machine learning model is trained, the optimal model is obtained through iteration, and the predicted volume of microplastics is output through the optimal model. Specifically, a volume prediction model based on PyCaret is constructed to output the predicted volume of microplastics. This includes: collecting microplastic samples with known volumes, obtaining their microscopic images and Raman data, extracting shape factors, shape parameters, and particle area, and establishing a machine learning model based on the above morphological feature parameters of each particle. The model uses particle area (A), shape parameter (i.e., aspect ratio (AR), and shape factor (i.e., roundness (C)) as input features, and the known measured volume (converted by weighing-density) as the training target. It iterates through 10 machine learning regression models (such as random forest, XGBoost, support vector regression, etc.) in the PyCaret library, performs 10-fold cross-validation (evaluation index R²>[0.8]), selects the prediction model constructed by the random forest regression algorithm as the optimal model, and performs regression prediction on unknown microplastic particles in batches through the optimal model to output the predicted volume of microplastic particles.

[0021] S4. Multidimensional feature integration output of microplastics, including microplastic sample particle ID, particle size, shape category, area A, aspect ratio AR, solidity S, roundness C, predicted volume V, and material.

[0022] Specifically, a structured data table was constructed using a Python script (Pandas library). Rows represent microplastic sample particle IDs, and columns include: particle size (equivalent circle diameter, μm), shape category, and area (μm²). 2 ), Aspect Ratio (AR), Solidity (S), Circularity (C), Predicted Volume (V, μm) 3 Materials, output as a CSV file.

[0023] The statistical report also includes: the abundance (number of particles) of microplastics, the distribution of the proportion of each material, the volume distribution and the shape distribution.

[0024] This invention also provides a rapid detection system for multidimensional features of microplastics applied to the above-mentioned detection method, comprising: The image detection module is used to identify microplastic particles in microscopic images and extract morphological feature parameters; The spectral analysis module is used to match microplastic materials based on Raman spectral data; The volume prediction module has a built-in machine learning model for predicting the volume of microplastics based on morphological characteristics. The data integration module is used to generate statistical reports on the multidimensional characteristics of microplastics.

[0025] Example 1 Rapid detection of single particles of microplastics in laboratory simulation This embodiment verifies the feasibility of the present invention for simultaneous detection of multidimensional features of single particles in discrete, low-concentration microplastic samples. The specific implementation process is as follows: (1) Laboratory-simulated microplastics (poly(dimethyl terephthalate) PET and nylon PA, with a particle size range of 500-2000 μm) were uniformly dispersed on a glass slide, covering a 5 mm × 5 mm detection area. Microscopic panoramic images of the samples were obtained using a low-power microscope (4× objective lens). Microplastic particles were identified by grayscale threshold segmentation (threshold = 130, range 0-255). Each image generated a particle coordinate position, where PET: (30021, 38804), PA: (35662, 18839).

[0026] The Raman spectrometer (excitation wavelength 785 nm, integration time 1-5 s) was automatically positioned to the coordinate point to collect Raman spectral data. The collected Raman spectral data was then matched with a preset spectral library (containing characteristic spectra of 30 common microplastics) based on the characteristic peak position and intensity similarity threshold. When the correlation coefficient threshold was ≥0.80, the material of the microplastic was determined.

[0027] (2) The acquired microscopic images are input into a pre-trained DeepLabV3 neural network. An edge detection algorithm is used to accurately segment the microplastic particles in the microscopic images. The model outputs a binary mask of the particles (e.g., ...). Figure 2 As shown in the figure, the model performance index mIOU (mean intersection-union ratio) reached 0.98. Using OpenCV to traverse the mask contour, the microplastic particles were identified and the morphological feature parameters of each particle were calculated and extracted (including: particle area (A), aspect ratio (AR), roundness (C), and solidity (S)). Based on the aspect ratio (AR) and roundness (C) thresholds, the particles were automatically classified, and it was determined that the shapes of PET and PA microplastic particles were fragmented.

[0028] (3) The extracted particle area (A), aspect ratio (AR), and roundness (C) features are input into the optimal model for volume prediction. The performance indicators of the model under 10-fold cross-validation are shown in Table 1 below. Its average R 2 The accuracy reached 0.81, demonstrating that the model has good prediction accuracy and stability, and the predicted PET particle volume is 1.25 × 10⁻⁶. 6 μm³, the volume of PA particles is 3.10 × 10⁻⁶. 6 μm 3 .

[0029] Table 1 Performance of the optimal volume regression model

[0030] (4) Multidimensional feature integration output of microplastics: A structured data table was constructed using a Python script (Pandas library), where rows represent microplastic particle IDs (ID_1, ID_2, ..., ID_n) (see Table 2 below for results), and columns include: particle size (equivalent circle diameter, μm), shape category, and area (μm²). 2 ), Aspect Ratio (AR), Solidity (S), Circularity (C), Predicted Volume (V, μm) 3 The results (see Table 3) are output as a CSV file, supporting subsequent statistical analysis.

[0031] Table 2 Multidimensional Feature Data Table

[0032] Table 3. Multidimensional Characteristic Data of Microplastics in Laboratory Samples

[0033] Application Example 1 Rapid detection of environmental microplastics This application example uses real-world samples to verify the ability of the method of this invention to perform rapid, simultaneous, multi-dimensional feature detection of multi-particle, multi-morphological microplastics in complex backgrounds. The specific implementation process is as follows: (1) Surface water samples from a city river were collected and filtered through a 2 μm stainless steel filter membrane. The filter membrane containing particulate matter was then placed on a glass slide for testing. A 5 mm × 5 mm panoramic microscopic image of the sample area was obtained using a low-power microscope (4× objective lens). Five suspicious targets were initially identified by grayscale threshold segmentation (threshold = 145), and a coordinate list was generated. A Raman spectrometer (excitation wavelength 785 nm, integration time 3 s) was automatically positioned at each coordinate point and spectral data was collected. The data was then matched with a preset spectral library, and a correlation coefficient threshold of ≥0.80 was set. Five of the samples were determined to be microplastics. The coordinates and material determination results of the five microplastic particles are as follows: The coordinates and material determination results of the 5 microplastic particles are as follows: Fiber 1-1 (coordinates: 12150, 22190): Polyethylene (PE), correlation coefficient 0.86 Particle 2-1 (coordinates: 16350, 22177): Styrene-butadiene rubber (SBR), correlation coefficient 0.96 Particle 2-2 (coordinates: 34210, 18906): Styrene-butadiene rubber (SBR), correlation coefficient 0.94 Fiber 2-3 (coordinates: 45000, 35093): Polyethylene (PE), correlation coefficient 0.82 Particles 2-4 (coordinates: 51000, 40301): polyethylene terephthalate (PET), correlation coefficient 0.95.

[0034] (2) Input the acquired microscopic images into a pre-trained DeepLabV3 neural network (backbone network ResNet-50, output stride = 16) to perform precise contour segmentation on the above 5 microplastic particles and output a binary mask (e.g. Figure 3 (As shown in Table 4). The model achieves an mIOU of 0.95 under complex backgrounds. OpenCV was used to traverse the mask contours and identify and calculate the shape characteristic parameters of each particle.

[0035] Based on AR and C threshold classification, fibers 1-1 and 2-3 were identified as fibrous, while particles 2-1, 2-2, and 2-4 were identified as granular (particles 2-1 and 2-4 were fragmented, and particle 2-2 was spherical), consistent with actual observations.

[0036] Table 4. Results of morphological feature extraction of microplastics from environmental samples

[0037] (3) Using the (particle) area (A), aspect ratio (AR), and roundness (C) extracted from Table 4 as input features, the regression model based on PyCaret, i.e. the optimal model, is called to predict the volume. The predicted volume results for each particle are as follows: Fiber 1-1 (PE): 20908 μm 3 Particle 2-1 (SBR): 100069 μm 3 Particle 2-2 (SBR): 140591 μm 3 Fiber 2-3 (PE): 17702 μm 3 Particle size 2-4 (PET): 88000 μm 3 .

[0038] (4) All the above morphological features are integrated by Python script (Pandas library) to generate a structured data table containing particle ID, particle size (equivalent circle diameter), shape category, area, aspect ratio, solidity, roundness, predicted volume and material. The results are shown in Table 5 and output as a CSV file to support subsequent environmental statistical analysis.

[0039] Table 5. Multidimensional Characteristic Data of Microplastics in Environmental Samples

[0040] This application example successfully verifies the application value of the method of the present invention in real environmental samples. In a single process, it simultaneously and rapidly and accurately detects the multidimensional characteristics (particle size, shape, volume, and material) of five different morphologies (two fibers and three particles) and different materials of microplastics, fully demonstrating the efficiency and comprehensiveness of the method. In summary, the high-throughput rapid detection method for multidimensional features of microplastics provided by this invention achieves the following: (1) significantly improved detection efficiency: the detection time for a single sample is shortened to 15 minutes; (2) comprehensive guarantee of detection accuracy: material identification accuracy ≥95%; morphological classification accuracy ≥92%; average error in volume prediction <40%; (3) multidimensional synchronous analysis capability: for the first time, multiple core morphological features are output synchronously in a single process, avoiding sample loss and data deviation caused by step-by-step detection. This provides a reliable tool for the study of microplastic pollution mechanisms and environmental monitoring, and therefore has broad application prospects.

[0041] This invention is not limited to the specific embodiments described above. Any modifications made by those skilled in the art based on the above concept without creative effort are within the scope of protection of this invention.

Claims

1. A high-throughput rapid detection method for multidimensional features of microplastics, characterized in that, Includes the following steps: S1. Obtain a microscopic image and corresponding coordinates of the microplastic sample, and obtain the Raman spectral data of the coordinate point. Based on the obtained Raman spectral data, determine the material of the microplastic by spectral matching. S2. Perform automated image segmentation processing on the microscopic images obtained in step S1 to identify microplastic particles and extract morphological feature parameters, including: Equivalent circle diameter: as particle size; Aspect ratio: The major axis / minor axis of the smallest bounding rectangle, used as a shape parameter; Circularity: 4πA / circumference 2 , and treat it as a shape factor; Particle area: Number of outline pixels × calibration coefficient; And the number of particles; Then, the shape of the microplastic particles is automatically determined based on the aspect ratio and roundness threshold. S3. Construct a volume prediction model based on PyCaret: Based on the morphological feature parameters of each particle, establish a machine learning model, wherein the model takes shape factor, particle size and particle area as input features, and the measured volume as the training target to train the machine learning model and output the predicted microplastic volume. S4. Microplastic multidimensional feature integration output, including microplastic particle ID, particle size, shape category, area A, aspect ratio AR, solidity S, roundness C, predicted volume V, and material.

2. The high-throughput rapid detection method for multidimensional features of microplastics according to claim 1, characterized in that, The specific steps of step S1 include: S11. Use a low-power microscope to obtain microscopic images of the sample, and identify microplastic particles by grayscale threshold segmentation to obtain the coordinate positions of the microplastic particles. S12. Automatically position the Raman spectrometer to the coordinate position and collect Raman spectral data; based on the Raman spectral data, determine the material of the microplastic by matching the spectral library; S13. Perform spectral matching between the obtained Raman spectral data and the preset spectral library based on the characteristic peak position and intensity similarity threshold to determine the material of the microplastic.

3. The high-throughput rapid detection method for multidimensional features of microplastics according to claim 1, characterized in that, Step S2 specifically includes: inputting the microscopic image obtained in step S1 into a pre-trained DeepLabV3 neural network to perform precise contour segmentation of the microplastic particles in the microscopic image and output a binary mask; then filtering out impurity interference through morphological operations, using OpenCV to traverse the mask contour, identifying the microplastic particles and calculating and extracting the morphological feature parameters of each particle, including the equivalent circle diameter of each particle as the particle size, the aspect ratio AR as the shape parameter, the roundness C as the shape factor, and the number of particles, and automatically classifying according to the aspect ratio AR and roundness C thresholds to determine the shape of the microplastic particles.

4. The high-throughput rapid detection method for multidimensional features of microplastics according to claim 1, characterized in that, In step S3, the machine learning model is a volume prediction model constructed using the random forest regression algorithm, and the training sample set includes measured volume data of microplastics of different materials and their corresponding shape factors, shape parameters, and particle areas.

5. The high-throughput rapid detection method for multidimensional features of microplastics according to claim 4, characterized in that, The machine learning model is trained as follows: collect microplastic samples of known volume, obtain their microscopic images and Raman data, extract shape factor, particle size, and particle area features, use the measured volume as a label, traverse the machine learning regression models in the Pycaret library, select the optimal model, and use it for volume prediction of microplastic particles.

6. The high-throughput rapid detection method for multidimensional features of microplastics according to claim 3, characterized in that, The principle for automatically classifying and determining the shape of microplastic particles based on aspect ratio (AR) and roundness threshold (C) is as follows: when AR > 5, the shape of the microplastic particles is fibrous; when C < 0.7 and AR ≤ 5, the shape of the microplastic particles is fragmented; when C < 0.7 and AR ≤ 5, the shape of the microplastic particles is spherical.

7. The high-throughput rapid detection method for multidimensional features of microplastics according to claim 1, characterized in that, The integrated output of the multidimensional features of microplastics also includes: microplastic abundance, distribution of the proportion of each material, volume distribution and shape proportion distribution.

Images