Plastic material discrimination method, discrimination device, and discrimination program
The method and device use visible light and near-infrared imaging with a learning model to enhance plastic material discrimination accuracy, addressing surface irregularities and cost issues, enabling efficient and accurate material identification.
Patent Information
- Application Number
- JP2023073180
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2023-04-27
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2043-04-27
AI Technical Summary
Existing plastic determination devices using near-infrared spectroscopy face challenges in achieving high discrimination accuracy due to surface irregularities and similarities in spectra of different plastics, requiring expensive sensors and lacking versatility for mass production.
A method and device utilizing visible light and near-infrared rays to capture images, employing a learning model to analyze hue and brightness histograms of these images, and a smartphone-based system for material discrimination.
Achieves highly accurate and stable plastic material discrimination, improving accuracy over time with a versatile and cost-effective solution suitable for mass production and portable use.
Smart Images

Figure 2024158193000001_ABST
Abstract
Description
[Technical field]
[0001] The present invention relates to a method, an apparatus and a program for identifying materials that are useful for separating plastic materials and recycling them during disposal. [Background technology]
[0002] Many methods and devices using near-infrared spectroscopy have been proposed as a means of identifying plastic materials.
[0003] Among these, the plastic determination devices described in Patent Documents 1 and 2 also use near-infrared spectroscopy, but in particular, the main device with a built-in central processing unit is made up of a tablet terminal, and the measuring device equipped with a spectroscopic sensor is also a handy type, so that anyone can easily carry the entire device anywhere and use it to determine the material of even plastics that are fixed in place and cannot be moved. In this respect, these are considered to be the publicly known determination devices that are closest to the present invention. [Prior art documents] [Patent documents]
[0004] [Patent Document 1] Patent No. 6947447 [Patent Document 2] Patent No. 7015579 Summary of the Invention [Problem to be solved by the invention]
[0005] However, the plastic determination devices disclosed in the above Patent Documents 1 and 2 only use near-infrared spectroscopy, and utilize the value of a reference spectrum (spectral spectrum) that is, so to speak, predetermined for each known material, to compare and collate the value of the spectral spectrum of an unknown material acquired by a spectroscopic sensor, and determine the material based on the most approximate value. Therefore, the determination result is easily influenced by, for example, the unevenness of the surface of the waste plastic, other roughness conditions, the size of the detection area, the performance of the spectroscopic sensor, and the like, and further, even plastics of different materials can have very similar spectra, making it difficult to achieve high determination accuracy (correct rate).
[0006] Furthermore, even if the above-mentioned plastic judgment device is used repeatedly and the number of times is increased, the accuracy of the judgment is not improved, and in order to improve the accuracy, it is necessary to adopt special and expensive near-infrared spectroscopic sensors and spectral cameras, which means that the device is inferior in versatility and mass productivity. [Means for solving the problem]
[0007] The present invention aims to fundamentally solve such problems, and in order to achieve this aim, claim 1 provides a method for identifying the material of plastics, which comprises irradiating a plastic of unknown material with visible light and near-infrared light, capturing the light reflected from the surface of the plastic with a camera, obtaining a visible light image and a near-infrared light image, and
[0008] measuring a hue histogram of the visible light image and a luminance histogram of the near-infrared light image,
[0009] The method is characterized in that the material of the plastic is identified using a learning model that a computer has machine-learned based on the features contained in each of the histograms.
[0010] In claim 2, a method for identifying the material of plastics is also described, which comprises cutting out and subdividing a certain area, which is a selected material identification target portion of plastic in a visible light image and a near-infrared light image, into a square grid, and measuring a hue histogram of the visible light image and a luminance histogram of the near-infrared light image for each section of the grid, and
[0011] The discrimination of the plastic material using the computer learning model is also performed for each section of the grid, and the discrimination results are tallied for all sections of the grid. The most frequent value obtained by the tabulation is used to discriminate the material as being the plastic.
[0012] Claim 3 also features a material discrimination method in which multiple near-infrared light images are obtained by irradiating a plastic of unknown material with multiple near-infrared light having different wavelengths and photographing the reflected light from the plastic.
[0013] In claim 4, a plastic material discrimination device is provided, which includes a measurement light irradiation device that irradiates a plastic of unknown material with visible light and near infrared light,
[0014] A camera captures the light reflected from the plastic surface during irradiation;
[0015] The computer includes at least an image analysis and processing unit, a learning unit, and a material discrimination unit.
[0016] The image analysis and processing unit of the computer measures the hue histogram of the visible light image and the luminance histogram of the near-infrared light image captured by the camera,
[0017] Similarly, the learning section of the computer performs machine learning of the features contained in both histograms based on a number of plastics whose materials are known in advance, and generates a learning model that serves as a material discrimination standard by inputting all the values of the hue histogram and brightness histogram.
[0018] Furthermore, the material discrimination section of the computer uses the learning model output from the learning machine to discriminate the material of unknown plastics.
[0019] In claim 5, the material discrimination device is also described, in which the computer is a smartphone or other mobile terminal having a camera function,
[0020] The measurement light irradiation device comprises an attachment or adapter to which the computer can be detachably attached and used,
[0021] The lower part of the measurement light irradiation device is pressed against the plastic surface and used as a lens barrel that defines the shooting space for the camera.
[0022] A white LED and at least one near-infrared LED are arranged in parallel on the barrel of the lens barrel.
[0023] The computer and the measurement light irradiation device are electrically connected so that the white LED and the near-infrared LED can be individually controlled to be turned on and off by an output signal from the computer.
[0024] In claim 6, the material discrimination device is also a smartphone with a bottom camera, and the upper part of the measurement light irradiation device, which is an attachment or adapter, is a smartphone receiving and supporting frame, and is shaped into an upward-open lip channel groove type that allows the smartphone to be inserted and removed from the front.
[0025] A transparent opening corresponding to the bottom camera of the smartphone is formed on the flat bottom surface of the groove of the lip channel groove type support frame, and the transparent opening is formed in a state of communication with the lens barrel which becomes the lower shooting space.
[0026] The measurement light irradiation device is characterized in that a white LED is used as the first light source, and two near-infrared LEDs with different wavelengths are used as the second and third light sources, which are attached in three tiers above, middle and bottom on the barrel surface of the lens barrel.
[0027] In claim 7, a plastic material discrimination program includes the steps of irradiating a plastic of unknown material with visible light and near-infrared light, and capturing the light reflected from the surface of the plastic with a camera to obtain a visible light image and a near-infrared image.
[0028] measuring the values of a hue histogram of the visible light image and a luminance histogram of the near-infrared light image in a certain region that is a selected material discrimination target portion of the plastic in each of the acquired images;
[0029] inputting all values of the hue histogram and the luminance histogram for the unknown plastic into a learning machine in which a computer has previously machine-learned the feature amounts contained in the hue histogram and the luminance histogram based on a plurality of different plastics whose materials are known;
[0030] The method is characterized in that the computer is caused to execute a step of determining the material of the unknown plastic using the learning model output from the learning machine and displaying the determined material.
[0031] In claim 8, the material discrimination program also includes the steps of irradiating a plastic of unknown material with visible light and near-infrared light, and capturing the light reflected from the surface of the plastic with a camera to obtain a visible light image and a near-infrared image.
[0032] a step of cutting out and subdividing a certain region, which is a selected material discrimination target portion of the plastic in each of the acquired images, into a square grid, and measuring each value of a hue histogram of the visible light image and a brightness histogram of the near-infrared light image for each section of the subdivided grid;
[0033] inputting all the values of the hue histogram and the luminance histogram measured in each section of the grid for the unknown plastic into a learning machine in which a computer has previously machine-learned the feature amounts contained in the hue histogram and the luminance histogram based on a plurality of different plastics whose materials are known;
[0034] The method is characterized in that the computer is caused to execute a step of outputting a material discrimination result for each grid section of the unknown plastic from the learning machine, and discriminating the material of the unknown plastic from the most frequent value obtained by aggregating the discrimination results output for each section as that for all sections, and displaying the discriminated material. Effect of the Invention
[0035] According to the configuration of claim 1, a visible light image and a near-infrared image are obtained by capturing the light reflected from the surface of the resin to be identified (a plastic of unknown material) when it is irradiated with visible light and near-infrared light, respectively, using a camera, and the features contained in the hue histogram of the visible light image and the brightness histogram of the near-infrared image, respectively, are input into a learning machine that has previously performed machine learning, and the material of the resin to be identified is identified using the characteristic shading distribution pattern of the color tone in these histograms and the characteristic light and dark (shade) distribution pattern that serves as the background, so to speak, as a learning model (teacher model) learned by the learning machine as the material identification standard.This method has the effect of stably obtaining highly accurate identification results, and by using the learning model as the next learning model and repeatedly updating it, the identification accuracy can be further improved.
[0036] In that case, particularly if the configuration of claim 2 is adopted, even if the resin to be identified is, for example, waste plastic and there is a risk that the reflected light from its surface cannot be captured stably due to unevenness or other roughness on the surface, by using the learning model generated by the learning machine, it is possible to always reliably detect even slight differences between plastics of different materials without overlooking them, and it is possible to achieve a 100% accuracy rate in identifying the material.
[0037] Furthermore, if the configuration of claim 3 is adopted, it is possible to train a learning machine to learn many of the features contained in the brightness histograms from multiple near-infrared light images, and this has the advantage of being able to use the acquired learning model to obtain highly accurate discrimination results.
[0038] According to the configuration of claim 4, a material determining device for using the above material determining method can have a simple configuration with the minimum necessary, and there is an effect that convenience and versatility in use can be obtained.
[0039] In that case, particularly if the configuration of claim 5 is adopted, the material discrimination device can be easily carried around, allowing anyone, anywhere, to discriminate the material of the resin being discriminated, which not only further improves the convenience of use but also helps achieve mass production economies.
[0040] Furthermore, if the configuration of claim 6 is adopted, a general smartphone can be used as the computer and the resin to be identified can be photographed using its bottom camera, so that it is sufficient to manufacture only the measurement light irradiation device that serves as an attachment or adapter for the smartphone, further improving mass productivity and widespread use.
[0041] According to the configuration of claim 7, by installing the plastic material discrimination program as a dedicated application on a smartphone, for example, anyone can instantly discriminate the material of the resin being discriminated in about 5 seconds.
[0042] In particular, according to the configuration of claim 8, the material discrimination program is executed by a computer, whereby the above-mentioned excellent effects associated with the configuration of claim 2 can be reliably achieved. [Brief description of the drawings]
[0043] [Figure 1] 1 is a functional block diagram of a material determining device according to an embodiment of the present invention. FIG. [Diagram 2] FIG. 2 is a perspective view showing a physical configuration of the material determining device. [Diagram 3] 3 is a front view showing an arrangement state of first to third light sources in the measurement light irradiating device extracted from FIG. 2. FIG. [Figure 4] FIG. 13 is an explanatory diagram of the process of cutting out a certain region of each image captured by a camera, which is a target portion for material discrimination, into a square grid. [Diagram 5] 1 is a graph illustrating one hue histogram and two brightness histograms for a polypropylene material. [Figure 6] 10 is a flowchart showing a material discrimination process for a plastic (a resin to be discriminated) whose material is unknown. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0044] A preferred embodiment of the present invention will be described in detail below with reference to the drawings. As can be seen from the functional block diagram of Figure 1, the plastic material discrimination device according to this embodiment comprises a device (A) that irradiates a plastic (hereinafter, also referred to as "resin to be discriminated") (10) whose material is unknown with measurement light, a camera (11) that captures the light reflected from the surface of the resin to be discriminated (10) that is the irradiated measurement light, and a computer (C) that measures the hue histogram and brightness histogram of the captured image, and discriminates the material of the resin to be discriminated (10) by using a learning model that serves as a material discrimination standard and that has been generated by a learning machine that has previously performed machine learning on plastics whose material is known.
[0045] In this regard, in the material discrimination device of the illustrated embodiment, as shown in Fig. 2, the computer (C) is embodied as a smartphone with a bottom camera (11), and the attachment or adapter serves as the measurement light irradiating device (A), to which the computer (hereinafter sometimes abbreviated as "smartphone") (C) can be detachably attached and used. The smartphone (C) and the separate measurement light irradiating device (A) are electrically connected via a control unit (control board) (12) by a USB cable (not shown), but the control unit (12) may be miniaturized and embedded inside the measurement light irradiating device (A).
[0046] The physical configuration of the material discrimination device shown in the figure consists simply of a smartphone (C) and a measurement light irradiation device (A) that serves as its attachment or adapter, so it can be carried anywhere and is conveniently used to discriminate the material of not only small, lightweight plastic products that can be moved freely, but also large, heavy plastic equipment that is fixed in place, and there is no need to prepare the resin to be discriminated (10) as a piece of a fixed size and shape.
[0047] In the illustrated embodiment, a smartphone with a bottom camera (11) is used as the computer (C) of the material discrimination device, but it is also possible to use a tablet terminal, a laptop computer, or other mobile communication terminal with a camera function instead. However, the camera (11) for capturing the reflected light may be built into the measurement light irradiation device (A), and the camera (11) and the image analysis / processing unit of the computer (C), which will be described later, may be connected so as to enable short-range wireless communication, for example, by Wi-Fi or preferably Bluetooth.
[0048] Among the main components of the above-mentioned plastic material discrimination device, the measurement light irradiation device (A) is formed or assembled into a single unit from an upper part that receives and supports the smartphone (computer) (C) as shown in Figures 1 and 2, and a lower part that defines the shooting space (S) for the camera (11).
[0049] That is, the upper part of the measurement light irradiation device (A) is shaped as a lip channel groove type that opens upward when viewed from the front as a receiving support frame (13) for the smartphone (C), and the smartphone (C) can be inserted and removed freely from the front (front side) along a pair of left and right facing slide guide grooves (14). (15) is a transparent opening formed in the flat groove bottom surface (13a) of the support frame (13) so that when the smartphone (C) is inserted, it corresponds to the bottom camera (11), and is connected to the shooting space (S) below.
[0050] On the other hand, the lower part of the measurement light irradiation device (A) is a square (preferably rectangular) or circular lens barrel (16) that defines the shooting space (S) for the camera (11), and the lower edge of its opening is pressed against the surface of the resin to be determined (10).
[0051] In this case, since the smartphone receiving support frame (13) forming the upper part of the measurement light irradiation device (A) is larger than the lens barrel (16) forming the lower part of the measurement light irradiation device (A), the user can grasp the left and right ends of the support frame (13) that protrude from the lens barrel (16) with both hands, and can reliably and conveniently press the lower edge of the opening of the lens barrel (16) against the resin (10) to be determined.
[0052] Although not shown in the figure, in order to prevent measurement light from leaking from the photographing space (S) of the camera (11), it is desirable to attach a sealing cover made of, for example, elastically deformable rubber, a flexible sheet, or some other material to the lower edge of the opening of the lens barrel (16).
[0053] As shown in Figures 1 and 3, on the body surface of the lens barrel (16) in the above-mentioned measurement light irradiation device (A), a pair of first to third light sources (17), (18a), and (18b) directed toward the shooting space (S) of the camera (11) are installed in parallel in three rows (top, middle, and bottom) at a required interval. These light sources are replaceable with new ones.
[0054] A white LED is used as the first light source (17) at the upper position, and is configured to irradiate white visible light onto the plastic (resin to be identified) (10) whose material is unknown.
[0055] An 850 nm near-infrared LED is used as the second light source (18a) in the middle position, and a 940 nm near-infrared LED is used as the third light source (18b) in the lower position, and these different near-infrared rays are similarly irradiated onto the resin to be evaluated (10).
[0056] The wavelengths used here (850nm and 940nm) are within the range of wavelengths that can be captured by the bottom camera (11) of the smartphone (C) and are preferable because they are relatively easy to obtain near-infrared LEDs, but there is no problem in using LEDs (light-emitting diodes) that belong to other near-infrared wavelength ranges (for example, 940nm and 960nm). It is sufficient to select a specific near-infrared wavelength that has a high correlation with the feature quantity to be output, which will be described later.
[0057] The smartphone (C) and the measurement light emitting device (A) are electrically connected so that the first to third light sources (17), (18a), and (18b) are all controlled to be turned on (lighted) and off (lighted off) by an output signal from the smartphone (C), and for example, one second after the first to third light sources (17), (18a), and (18b) are turned on, an image is taken by the camera (11), and 0.5 seconds after that, the light is turned off. (19) is an operation unit such as a push button that turns on and off the power of the first to third light sources (17), (18a), and (18b).
[0058] The relative arrangement of the first to third light sources (17), (18a), and (18b) in the three upper, middle, and lower tiers, the order in which they are turned on and off, and the order in which they are photographed by the camera (11) can be freely set without being limited to the numerical order of the first to third light sources (17), (18a), and (18b). A mobile battery that serves as the power source for the first to third light sources (17), (18a), and (18b) is not shown in the figure.
[0059] Next, the bottom camera (11) of the smartphone (computer) (C) captures (a total of three images) the light reflected from the surface of the resin to be identified (10) being irradiated with visible light from the white LED of the first light source (17), and the light reflected from the surface of the resin to be identified (10) being irradiated with near-infrared light from near-infrared LEDs of different wavelengths of the second and third light sources (18a) (18b), thereby obtaining one visible light image and two near-infrared images of different wavelengths, and the three images are input into the image analysis and processing unit of the computer (smartphone) (C), which will be described later.
[0060] Furthermore, as shown in the block diagram of Figure 1, the computer (smartphone) (C) has an image analysis / processing unit (20), a learning unit (21), a material discrimination unit (22), a memory unit (23), and a display unit (24), and a processor (e.g., a CPU) not shown in the figure, which constitutes the physical configuration, executes an operating system (OS) and application software (material discrimination program) (hereinafter sometimes abbreviated as a "dedicated app"), thereby enabling the material of the plastic (resin to be discriminated) (10) of unknown material to be discriminated.
[0061] That is, as described above, the image analysis and processing unit (20) of the computer (C) is in a connected state capable of wireless communication with the camera (11), and the one visible light image and two near-infrared light images captured by the camera (11) are output from the camera (11) to the image analysis and processing unit (20) of the computer (C).
[0062] In the image analysis and processing unit (20), as shown in Fig. 4, from each of the three images (P) taken by the camera (11), a certain region (size / area) (Z) that is the part to be identified as the material of the plastic (resin to be identified) (10) selected (adjusted) by the user is extracted, and the extracted certain region (Z) is divided into sections (25) of the same square grid (G) to generate a hue histogram of the above visible light image and a luminance histogram of two near-infrared light images with different wavelengths as shown in Fig. 5 (a), (b), and (c), and each value is measured. Fig. 5 shows three histograms for polypropylene.
[0063] Furthermore, the learning unit (21) of the computer (C) classifies a number of materials (six types in the illustrated embodiment) whose materials are known in advance using easy-to-understand binary numerical values (binarization) as shown in Table 1, and for each of the classified types, the learning machine uses a neural network to extract feature amounts contained in the hue histogram and both brightness histograms, and performs machine learning using deep learning to create a learning model (reference model / teacher model) (26) that serves as the material discrimination standard for the resin to be identified (10), and stores this in the memory unit (23) just like the application software (material discrimination program).
[0064] [Table 1]
[0065] In other words, in the analysis and processing by the computer (C) of the image taken by the camera (11), a simple binary number different for each material as shown in Table 1 is assigned as an identifier, so to speak, for multiple plastics, which are classified (for example, six types), and the identifier is input to a learning machine, which is trained by machine learning, and the learning model and the material discrimination results based on this are output as materials automatically converted from the identifier (binary number), thereby identifying and displaying the discriminated material. Table 1 is the conversion table for this purpose.
[0066] The material discrimination section (22) of the computer (C) inputs all the values of the hue histogram of the visible light image and the brightness histogram of both near-infrared light images acquired in the image analysis and processing section (20) for each section (25) of the grid (G) into a learning machine, and using the learning model stored in the memory section (23), it discriminates the material of the plastic (resin to be discriminated) (10) whose material is unknown, acting as an artificial intelligence (AI), and outputs the discrimination result for each section (25) of the grid (G) as one type of material classified in advance by a binary number as shown in Table 1. For example, if the display inside the computer is "1,0,0,0,0,0,0", it is converted to material: polypropylene (PP) and output.
[0067] Furthermore, the material discrimination unit (22) determines the material of the resin (10) to be discriminated from the most frequent value obtained by totaling the discrimination results for each section (25) for all sections of the grid (G), and displays the determined material on the display unit (24) of the computer (C).
[0068] The plastic material discrimination device according to the illustrated embodiment has the above-mentioned configuration, and the material discrimination process (material discrimination method) performed by this device will be described in detail below.
[0069] That is, when using the material discrimination device to discriminate the material of a plastic (a resin to be discriminated) (10) whose material is unknown, the lower edge of the opening of the lens barrel (16) of the measurement light irradiating device (A) is pressed firmly against the surface of the resin to be discriminated (10) in a state in which the measurement light (visible light and near-infrared light) does not leak out, while the first to third light sources (17), (18a), and (18b) are turned on by manually operating the operation unit (19) of the irradiating device (A), and a computer (smartphone) (e.g., a Google Pixel 5) (C) A dedicated application (material discrimination program) is started, and as shown in the flowchart of FIG. 6, first the white LED of the first light source (17) is turned on (Step S1), and while the visible light is being irradiated onto the resin to be discriminated (10) for a fixed period of time (e.g., 1 second), the light reflected from the surface of the resin to be discriminated (10) is photographed by the bottom camera (11) of the smartphone (C) (Step S2), thereby obtaining one visible light image.
[0070] After a certain time (e.g., 0.5 seconds) has elapsed after the photographing, the first light source (17) is turned off, and then the second light source (18a), an 850 nm near-infrared LED, is turned on (step S3). While irradiating the near-infrared light onto the resin to be identified (10) for the same certain time, the light reflected from the resin to be identified (10) is photographed by the camera (11) of the smartphone (C) (step S4), and a first near-infrared light image is obtained.
[0071] Furthermore, when the same fixed time has elapsed after the photographing, the second light source (18a) is also turned off, so the 940 nm near-infrared LED, which is the third light source (18b), is turned on (step S5), and while the near-infrared light is being irradiated onto the resin (10) to be identified for the same fixed time, the reflected light from the resin (10) is photographed by the camera (11) of the smartphone (C) (step S6), and a second near-infrared light image is obtained. When the same fixed time has elapsed after the photographing, the third light source (18b) is turned off.
[0072] As described above, the order of irradiating the measurement light from the first to third light sources (17), (18a), and (18b) and the order of capturing images by the camera (11) are not limited to the order shown in the flowchart of FIG. 6, and may be freely set.
[0073] Therefore, from each image (P) of one visible light image and two near-infrared light images taken by the camera (11), only a certain region (size / area) (Z) that is the part on the surface of the resin to be identified (10) that is the material identification target portion selected (adjusted) by the user is extracted (step S7), as shown in Fig. 4. The purpose of this is to cut out the central portion, so to speak, which is less susceptible to external light and other adverse effects and shows the surface condition well.
[0074] Moreover, the extracted fixed area (Z) is cut out and subdivided into the same square grid (G) for each of the three images (P) (step S8). If the surface of the resin to be identified (10) is uneven or has other irregularities, this may cause reflected light to be generated differently from that of a flat area, so the purpose of forming many areas capable of capturing the expected (learned) reflected light is to form the fixed area.
[0075] In the illustrated embodiment, the size of the square grid (G) is, for example, 256 x 256, and it is divided into a total of 25 sections, but in step S9 following step S8, as shown in Fig. 5 (a), (b), and (c) which shows an example of a graph of polypropylene material, the hue histogram (a) of the visible light image, the brightness histogram (b) of the 850 nm near-infrared light image, and the brightness histogram (c) of the 940 nm near-infrared light image are calculated (converted) for each section (25) of the grid (G) to obtain each value. In the graphs of Fig. 5 (a), (b), and (c), the vertical axis indicates the number of pixels, and the horizontal axis indicates the size (range: 0 to 255).
[0076] That is, in the graph showing the histograms in FIG. 5, one hue histogram value and two brightness histogram values each have 256 values (256 gradations), for a total of 768 values.
[0077] Then, all of the values of the hue histogram and both luminance histograms obtained for each section (25) of the grid (G) (a total of 768 values as exemplified above) are input into a learning machine in a learning unit (21) in a computer (smartphone) (C) for machine learning (step S10).
[0078] Then, the learning machine outputs the material discrimination result for each section (25) in the grid (G) as a material that is automatically converted from the binary number into one of the materials previously classified as shown in Table 1 (conversion table) (step S11).
[0079] In this case, it is presumed that the learning model (teacher model) (26) that serves as the material discrimination standard learned by machine learning in the learning machine comprehensively learns the waveforms of the histograms in the three input images (the hue histogram of the visible light image and the brightness histograms of two near-infrared images with different wavelengths), the positions and sizes (height and width of the spread of the tail) of the uneven parts, and the correlations when compared with the waveforms of other histograms, and thereby discriminates the material.
[0080] Therefore, the discrimination results for each section (25) in the grid (G) are subsequently compiled as those for all sections of the grid (G) as shown in Table 2, and the most frequent value obtained by the compilation is used to determine the material as the resin to be discriminated (10). The final determined material is then displayed on the display unit (24) of the computer (smartphone) (C) (step S12).
[0081] [Table 2]
[0082] In other words, if the tabulated results are as shown in Table 2, for example, then since “1,0,0,0,0,0” polypropylene is the most frequent value, the material of the resin to be identified (10) is determined to be polypropylene (PP) and this is displayed.
[0083] As is clear from the flowchart of FIG. 6, the method for identifying the material of plastic according to the illustrated embodiment includes steps (S1 to S6) of irradiating a plastic (10) of unknown material with visible light and two near-infrared rays of different wavelengths, and capturing the light reflected from the surface of the plastic (10) with a camera (11) to obtain one visible light image and two near-infrared images;
[0084] a step (S7 to S9) of cutting out and subdividing a certain region (Z) in each of the three acquired images (P) which is a target portion for material discrimination of the plastic (10) into a square grid (G), and measuring each value of the hue histogram of the visible light image and the brightness histogram of both near-infrared light images for each section (25) of the subdivided grid (G);
[0085] A step (S10) of inputting all the values of the hue histogram and both brightness histograms measured in each section (25) of the grid (G) of the unknown plastic (10) into a learning machine that has machine-learned the features of the hue histogram and both brightness histograms by performing the same steps as the steps described above for each type of plastic that is classified into numerical values in advance and classifying the plastics of known materials;
[0086] a step (S11) of outputting from the learning device a material discrimination result for each section (25) of the grid (G) in the unknown plastic (10) as a material corresponding to each of the numerical values classified in advance;
[0087] The learning machine outputs the discrimination results for each section (25) of the grid (G) and tallying them up for all sections to determine the most frequent value, which is then used to determine that the material is the unknown plastic (10) (step S12).
[0088] The material discrimination device used for this purpose is a measurement light irradiation device (A) that irradiates a plastic (10) of unknown material with visible light and two near-infrared rays with different wavelengths,
[0089] a camera (11) for photographing the light reflected from the surface of the plastic (10) on which the measurement light is irradiated;
[0090] The system comprises a computer (C) having at least an image analysis and processing unit (20), a learning unit (21) and a material discrimination unit (22),
[0091] The image analysis and processing section (20) of the computer (C) subdivides each of the one visible light image and the two near-infrared light images captured by the camera (11) into a square grid (G), and measures the values of the hue histogram of the visible light image and the luminance histograms of both near-infrared light images for each section (25) of the subdivided grid (G);
[0092] Similarly, in the learning section (21) of the computer (C), a learning machine that performs machine learning based on feature amounts contained in the hue histogram of the visible light image and the luminance histograms of both near-infrared light images for each section (25) of the grid (G) based on each type of quantified (binarized) classification of multiple plastics whose materials are known in advance generates a learning model (26) that serves as a material discrimination standard by inputting all the values of the hue histogram and both luminance histograms,
[0093] The material discrimination section (22) of the computer (C) discriminates the material of the unknown plastic (10) for each section (25) of the grid (G) using a learning model (26) output from the learning machine, and determines that the material is the unknown plastic (10) based on the most frequent value obtained by aggregating the discrimination results for all sections.
[0094] According to the above-described configuration of the illustrated embodiment, one visible light image and two near-infrared images are obtained by capturing, with a camera (11), light reflected from the resin to be identified (plastic of unknown material) (10) that is irradiated with visible light and two near-infrared light rays having different wavelengths. The feature amounts contained in the hue histogram of the visible light image and the brightness histograms of both near-infrared light images are input to a learning machine that has previously performed machine learning. The learning model (teacher model) (26) that the learning machine has learned is used as a material discrimination standard to discriminate the material of the resin to be identified (10). This has the effect of stably obtaining highly accurate discrimination results.
[0095] In other words, even if the characteristics of near-infrared reflected light change depending on the color of the plastic material, the material can be reliably identified by measuring the histogram (hue histogram) that represents the color of the material itself and the luminance histogram of near-infrared reflected light and inputting them into the computer learning machine. The learning model output from the learning machine is used as the next learning model, and by repeatedly updating it, the accuracy of material identification can be further improved.
[0096] In this case, two near-infrared images are obtained by irradiating near-infrared light of different wavelengths, and for both the brightness histograms of these near-infrared images and the hue histograms of the visible light images, a certain area (Z) of the part to be identified in each captured image (P) is divided into a square grid (G) and each value of the histogram is measured for each of a plurality of small sections (25). Therefore, even if there is a risk that the reflected light from the surface of the resin to be identified (10) cannot be captured stably due to unevenness or other roughness on the surface, combined with the use of the learning model (26) generated by the learning machine, even slight differences between plastics of different materials will not be overlooked and can be reliably found at all times, making it possible to achieve a 100% accuracy rate in identifying the material.
[0097] However, as is clear from steps S8 to S11 in the flowchart of FIG. 6, in the above-mentioned material discrimination method and apparatus, a certain region (Z) which is the target portion for material discrimination in each image (P) captured by the camera (11) is first divided into a square grid (G) and the values of the hue histogram and the brightness histogram are obtained for each of the finer sections (25). However, if the surface of the resin to be discriminated (10) is kept smoothly flat and without any problems, and even if the surface is uneven or has other irregularities, if the user can select (adjust) the certain region (Z) which is the target portion for material discrimination while avoiding the irregular parts, then there is no need to subdivide the region into the grid (G). Step S8 in FIG. 6 can be omitted, and the values of the hue histogram and the brightness histogram can be measured for the entire certain region (Z), and processing steps equivalent to steps S9 to S11 can be carried out sequentially. In the process of step S11, the learning machine outputs the final material discrimination result of the resin to be discriminated (10), and the material is displayed.
[0098] Furthermore, as suggested by steps S3 to S6 in the flowchart of FIG. 6, two near-infrared light images and therefore two brightness histograms are obtained by irradiating near-infrared light of different wavelengths. However, if the brightness histogram of the near-infrared light image is taken into consideration together with the hue histogram of the visible light image in determining the material of the resin to be identified (10), it is possible to only obtain one near-infrared light image and therefore one brightness histogram by irradiating the resin using one near-infrared LED as a light source. [Explanation of symbols]
[0099] (10) Plastic of unknown material (resin to be identified) (11) Camera (bottom camera) (12) Control section (13) Smartphone holder support frame (13a)... Groove bottom surface (14) Slide guide groove (15)...Transmission port (16) Telescope tube (17)...1st light source (18a)...Second light source (18b)...Third light source (19)...Operation section (20) Image analysis and processing unit (21) Learning Department (22)...Material discrimination section (23)...Storage section (24)...Display section (25) 1 plot (26) Learning Model (A) Measurement light irradiation device (C) Computer (smartphone) (G) Grid (P)···Image (S)...Shooting space (Z)...Constant area
Claims
1. A plastic sample of unknown material is irradiated with visible light and near-infrared light, and the light reflected from the surface of the plastic is photographed with a camera to obtain a visible light image and a near-infrared light image. measuring a hue histogram of the visible light image and a luminance histogram of the near-infrared light image, A method for identifying the material of plastics, characterized in that a computer identifies the material of the plastic using a learning model that has been machine-learned based on the features contained in each of the histograms.
2. A certain area, which is a part of the plastic material to be identified in the visible light image and the near-infrared light image, is cut out and subdivided into a square grid, and a hue histogram of the visible light image and a brightness histogram of the near-infrared light image are measured for each section of the grid. The method for identifying plastic materials as described in claim 1, characterized in that the discrimination of the plastic material using a computer learning model is also performed for each section of the grid, and the discrimination results are compiled for all sections of the grid, and the plastic material is identified based on the most frequent value obtained by the compilation.
3. A method for identifying the material of plastic as described in claim 1 or 2, characterized in that a plurality of near-infrared images are obtained by irradiating a plastic of unknown material with multiple near-infrared rays having different wavelengths and photographing the reflected light from the plastic.
4. A measurement light irradiation device that irradiates a plastic of unknown material with visible light and near infrared light, A camera captures the light reflected from the plastic surface during irradiation; The computer includes at least an image analysis and processing unit, a learning unit, and a material discrimination unit. The image analysis and processing unit of the computer measures the hue histogram of the visible light image and the luminance histogram of the near-infrared light image captured by the camera, Similarly, the learning section of the computer performs machine learning of the features contained in both histograms based on a number of plastics whose materials are known in advance, and generates a learning model that serves as a material discrimination standard by inputting all the values of the hue histogram and brightness histogram. Furthermore, the material discrimination unit of the computer uses the learning model output from the learning machine to discriminate the material of unknown plastics.
5. While computers consist of smartphones and other mobile devices with cameras, The measurement light irradiation device comprises an attachment or adapter to which the computer can be detachably attached and used, The lower part of the measurement light irradiation device is pressed against the plastic surface and used as a lens barrel that defines the shooting space for the camera. A white LED and at least one near-infrared LED are arranged in parallel on the barrel surface of the lens barrel, 5. The plastic material discrimination device according to claim 4, wherein the computer and the measurement light irradiation device are electrically connected so that the white LED and the near-infrared LED can be turned on and off by output signals from the computer.
6. The computer is a smartphone with a bottom camera, and the upper part of the measurement light irradiation device, which is an attachment or adapter, is a smartphone receiving and supporting frame, and is shaped into an upward open lip channel groove type that allows the smartphone to be inserted and removed from the front, A transparent opening corresponding to the bottom camera of the smartphone is formed on the flat bottom surface of the groove of the lip channel groove type support frame, and the transparent opening is formed in a state of communication with the lens barrel which becomes the lower shooting space. The plastic material discrimination device according to claim 5, characterized in that a white LED is mounted on the barrel surface of the measurement light irradiation device as the first light source and two near-infrared LEDs with different wavelengths are mounted in three rows above, middle and bottom as the second and third light sources.
7. A step of irradiating a plastic of unknown material with visible light and near-infrared light, and capturing the light reflected from the surface of the plastic with a camera to obtain a visible light image and a near-infrared light image; measuring the values of a hue histogram of the visible light image and a luminance histogram of the near-infrared light image in a certain region that is a selected material discrimination target portion of the plastic in each of the acquired images; inputting all values of the hue histogram and the luminance histogram for the unknown plastic into a learning machine in which a computer has previously machine-learned the feature amounts contained in the hue histogram and the luminance histogram based on a plurality of different plastics whose materials are known; A plastic material discrimination program for causing a computer to execute a step of discriminating the material of the unknown plastic using the learning model output from the learning machine and displaying the discriminated material.
8. A step of irradiating a plastic of unknown material with visible light and near-infrared light, and capturing the light reflected from the surface of the plastic with a camera to obtain a visible light image and a near-infrared light image; a step of cutting out and subdividing a certain region, which is a selected material discrimination target portion of the plastic in each of the acquired images, into a square grid, and measuring each value of a hue histogram of the visible light image and a luminance histogram of the near-infrared light image for each section of the subdivided grid; inputting all the values of the hue histogram and the luminance histogram measured for each section of the grid for the unknown plastic into a learning machine in which a computer has previously machine-learned the feature amounts contained in the hue histogram and the luminance histogram based on a plurality of different plastics whose materials are known; A plastic material discrimination program for causing a computer to execute the steps of outputting material discrimination results for each grid section of the unknown plastic from the learning machine, discriminating the material of the unknown plastic from the most frequent value obtained by aggregating the discrimination results output for each section as that for all sections, and displaying the discriminated material.
Citation Information
Patent Citations
Plastic determination device and plastic determination program
JP6947447B1
Plastic determination device and plastic determination program
JP7015579B1
