TWI672639B - Object recognition system and method using simulated object images - Google Patents

Abstract

The present invention provides an object recognition method using a simulated object image, the method comprising: (A) acquiring an image collection of an object including one or more object images and a background image set including a plurality of background images; (B) The object image collection and the background image collection generate a simulated object image collection including a plurality of simulated object images; (C) training a test object identification model according to the simulated object image set; and (D) is to be tested by a test object A to-be-measured image obtained by the scene is input into the object identification model to obtain an object identification result.

Description

Object recognition system using simulated object image and method thereof

The present invention relates to object recognition, and more particularly to an object recognition system using a simulated object image and a method thereof.

The training of the identification model is based on a large amount of annotation data, and the amount of data and the quality of the data affect the recognition rate of the training model. For some tasks or areas, this data can be collected over a long period of time to help solve problems in the field. Therefore, it takes time to collect data and classify and label it before training the model.

In the case of an identification system, the recognition rate depends on whether there are enough data samples, and the higher the diversity of the samples, the more the problems encountered in each field can be overcome. Therefore, a good identification model will take a lot of time to collect data and annotation data. In addition, when the specific field identification rate cannot reach the standard, the data of the field can be collected, and targeted training and adjustment can be applied to improve the recognition rate of the field. However, it also led to an increase in overall construction time and an increase in initial construction costs. On the other hand, in areas where private protection is more rigorous, it faces the dilemma of obtaining large amounts of data, and more resources must be spent on collecting data.

The present invention provides an object recognition system using a simulated object image and a method thereof to solve the problems encountered in the conventional identification system.

The present invention provides an object recognition method using a simulated object image, the method comprising: (A) acquiring an image collection of an object including one or more object images and a background image set including a plurality of background images; (B) The object image collection and the background image collection generate a simulated object image collection including a plurality of simulated object images; (C) training a test object identification model according to the simulated object image set; and (D) is to be tested by a test object A to-be-measured image obtained by the scene is input into the object identification model to obtain an object identification result.

The invention further provides an object recognition system using a simulated object image, comprising: a non-volatile memory for storing an object recognition program; and an operation unit for executing the object recognition program to perform the following steps: obtaining a collection of object images of a plurality of object images and a background image set including a plurality of background images; generating, according to the object image set and the background image set, a set of simulated object images including a plurality of simulated object images; The set is used to train a test object identification model; and a to-be-tested image obtained by a scene to be tested is input into the object identification model to obtain an object recognition result.

The following description sets forth various embodiments of the invention. The following description sets forth the basic concepts of the invention and is not intended to limit the invention. The scope of the actual invention shall be defined in accordance with the scope of the patent application.

1 is a block diagram showing an object recognition system in accordance with an embodiment of the present invention.

In an embodiment, the object recognition system 100 can be implemented in an electronic device, such as a personal computer, a server, or a portable device. The object recognition system 100 includes an operation unit 110, an image capture device 120, a storage unit 130, and a display 150.

The computing unit 110 can be implemented in various manners, such as a dedicated hardware circuit or a general-purpose hardware (for example, a single processor, a multi-processor with parallel processing capability, a graphics processor, or other processor with computing power), and When the code or software related to each model and process of the present invention is executed, the functions described later are provided. The image capturing device 120 is, for example, a camera for capturing a to-be-tested image for a scene to be tested.

The storage unit 130 includes a volatile memory 131 and a non-volatile memory 132. The non-volatile memory 132 is used to store a database of various image collections, data required in the object identification process, and various code codes, such as various algorithms and/or object recognition models. The non-volatile memory 132 can be, for example, a hard disk drive, a solid-state disk, a flash memory, or a read-only memory (read- Only memory), but the invention is not limited thereto. The volatile memory 131 can be a random access memory such as a static random access memory (SRAM) or a dynamic random access memory (DRAM), but the invention is not limited thereto. The volatile memory 131 can temporarily store intermediate data and images in the object identification process, for example.

In one embodiment, the non-volatile memory 132 stores an object recognition program 133, and the operation unit 110 reads the object recognition program 133 from the non-volatile memory 132 to the volatile memory 131 and executes the object. The identification program 133 includes a code for an object identification method.

The display unit 150 can be a display panel (for example, a thin film liquid crystal display panel, an organic light emitting diode panel, or other display capable panel) for displaying input characters, numbers, symbols, dragging a mouse's movement track or an application. The user interface provided is provided for viewing by the user. The object recognition system 100 further includes an input device (not shown) for the user to perform a corresponding operation, such as a mouse, a stylus, or a keyboard, but the present invention is not limited thereto.

In one embodiment, the non-volatile memory 132 further includes a first database 135, a second database 136, a third database 137, a fourth database 138, a fifth database 139, and a The sixth database 140 and a sample identification model 141. For example, the first database 135 stores a plurality of object scene images. Each object scene image may include, for example, one or more types of objects, wherein the objects may be, for example, text (eg, A to Z, 0 to 9, or other fonts), human body, license plate, components, and logos, etc., but The invention is not limited to this.

The second database 136 stores a plurality of background images, such as a background image collection. The background image may be, for example, a real background image of any real scene obtained under different shooting conditions, and is not limited to the background image of the scene to be tested, and may not include the object to be tested. In some embodiments, the background image further includes a virtual background image simulated by computer vision technology.

The third database 137 stores a plurality of object images, for example, an object image collection, wherein each object image is obtained by, for example, a plurality of object scene images in the first database 135. The fourth database 138 stores a plurality of simulated object images, for example, a simulated object image collection.

The computing unit 110 generates the simulated object image set in the fourth database 138 according to the object image set in the third database 137 and the background image set in the second database 136, the details of which are detailed in Rear.

The 2A-2M diagram shows a schematic diagram of different images used in the object recognition process in accordance with an embodiment of the present invention. Please also refer to Figure 1 and Figure 2A-2M. For convenience of explanation, the object to be tested used in the embodiment to be described later is a license plate.

The image of each object scene stored in the first database 135 can be, for example, a real license plate image, for example, including all license plate characters (for example, A to Z, 0 to 9, or other fonts), as shown in FIG. 2A. For example, the operation unit 110 can perform image capture processing on each object scene image to obtain an image of each character in the license plate (ie, an object image), as shown in FIG. 2B. The computing unit 110 uses optical character recognition technology or other object recognition technology to obtain all license plate characters, and each license plate character is a separate object image, as shown in FIG. 2C, 10 of which are The digital object image, and the 26-letter object image, wherein the object images of all the license plate characters can be stored, for example, in the third database 137.

Next, the computing unit 110 composes one or more training objects from one or more object images according to a predetermined rule. Since the license plate is taken as an example in this embodiment, the predetermined rule is a rule for the license plate, including, for example, the license plate length and width, the font spacing, the character limit, the character layout, the font color, the license plate color, the screw hole size and the position, and the like. Wait. Figure 2D shows the rules for the development of license plates for cars (self-use passenger cars), but the invention is not limited to automobile license plates, and license plates for other vehicle types can also be used, such as large heavy locomotives, ordinary heavy locomotives, buses, large Trucks and so on. That is, the license plate of different vehicle types has a corresponding license plate making rule, and the operation unit 110 can use the different combinations of the object images of the license plate characters according to the selected license plate to generate one or more training objects (for example, a simulated license plate). Image), as shown in Figure 2E. It should be noted that the simulated license plate image in Figure 2E is composed of object images of different license plate characters in the third database 137, and the simulated license plate is not added with noise, blur, type change or real. Various image features of the scene.

The computing unit 110 then performs the first image processing to add one or more object image features and one or more background image features to the simulated license plate image (ie, the training object). For example, the object image feature can be, for example, an object to be tested in a real scene, which is visually affected by the influence of the environment. Object image features include, for example, blurriness, scratches or stains, shadows, shadings, overexposures, distortions, and color aberrations. However, the invention is not limited to this. Figure 2F shows a schematic diagram of a license plate containing different image features. Because the object image features and the background image features include a plurality of different types of image features, the computing unit 110 can perform the first image processing to add one or more object image features to each training object (eg, a simulated license plate image) to generate one or A plurality of simulated objects to be tested (processed simulated license plate images), for example, the 2H-1 to 2H-6 images show that the simulated license plate images in FIG. 2E are respectively added with scratches, chromatic aberrations, shadows, blurs, noises, The simulated object to be tested generated after the image distortion of the object such as deformation and chromatic aberration. It should be noted that the present invention is not limited to adding only one of the object images to each training object (for example, a simulated license plate image).

Background image features such as noise generated by images captured in real scenes, and background image features may also be referred to as environmental noise features. Background image features include, for example, blurring, scratches or stains, shadows, noise, shadowing, overexposure, distortion, and chromatic aberration, but the invention is not limited thereto. The 2G graph shows a schematic representation of a real scene containing different background image features. The details of the object image features and background image features will be detailed later.

In some embodiments, the computing unit 110 may perform a first image processing to add one or more object image features and one or more background image features to each training object (eg, a simulated license plate image) to generate one or more simulations. Object to be tested. For example, in addition to the image features of the object that may appear on the license plate, the license plate image may also receive background noise images in the real scene, so the computing unit 110 may also feature one or more object images and one or more The background image is specially added to each of the training objects to generate one or more simulated objects to be tested.

In one embodiment, the background image stored in the background image set of the second database 136 is shown, for example, in FIG. It should be noted that the background image in FIG. 2I may not include a license plate.

Then, the computing unit 110 randomly selects a background image from the background image set stored by the second database 136, wherein the selected background image may be, for example, all or part of one of the real background images in the background image set ( For example, the region of interest) is shown in Figures 2J-1 and 2J-2, respectively. Assuming that the background image of the region of interest of the 2J-2 image is used (for example, the first background image), the computing unit 110 performs a second image processing to add one or more background image features to the first background image to generate A simulated background image. For example, the operation unit 110 may add one or more background image features such as blur, scratch or stain, shadow, noise, shadow, overexposure, deformation, etc. to the first background image, so that the first The background image scene can be incorporated into different image features that were not originally captured, so a smaller number of background images can be used to achieve the image effect of the background environment under different shooting conditions.

In the foregoing embodiment, the computing unit 110 may perform the first image processing to add one or more object image features and one or more background image features to each training object (eg, a simulated license plate image) to generate one or more simulations. The object to be tested is subjected to a second image processing to add one or more background image features to the first background image to generate a simulated background image. Because the simulated object to be tested is generated by adding one or more object image features to the part of the license plate, and the simulated background image is generated by adding one or more background image features to the first background image, but simulating the object to be tested and There may be no correlation between simulated background images. Therefore, the arithmetic unit 110 performs an image synthesis process to add the simulated object to be simulated to generate a simulated synthetic image, as shown in FIG. 2K.

For example, the image synthesizing process can adjust the simulated object to be measured to an appropriate image size and paste it at any position in the simulated background image (for example, in a predetermined range in the simulated background image), and perform the pairing. The simulated object to be tested, which is a simulated background image, is subjected to edge smoothing to generate a simulated synthetic image. It should be noted that the simulated object to be tested in the simulated background image itself does not have the image characteristics of the simulated scene in the simulated background image. Therefore, the computing unit 110 further performs second image processing to add one or more background image features to the simulated synthetic image, and generates a simulated object image, wherein the process is to enhance the consistency between the simulated object and the background to generate the image. Simulated object image for training. The 2L-1~2L-4 images respectively show the results of adding image, such as blur, interference, salt and pepper noise, and Gaussian noise, to the simulated synthetic image. The simulated object image shown in FIG. 2M is, for example, a result obtained by combining different background image characteristics in the second L-1 to 2L-4. In the above process of the present invention, overlaying the simulated object to be covered on any background image can improve the complexity of the background of the license plate and help to enhance the effect of the subsequent object recognition model training.

The computing unit 110 can select different combinations of object image features and background image features and select different real background images, and repeat the processes in the foregoing embodiments to generate different simulated object images. Therefore, the computing unit 110 can obtain a plurality of simulated object images to form a simulated object image set, and store the simulated object image in the fourth database 138.

Next, the computing unit 110 trains a DUT identification model 141 according to the simulated object image set in the fourth database 138. For example, the operation unit 110 may use a support vector machine (SVM), a convolutional neural network, a deep neural network, or the like to train the object identification model. 141, but the invention is not limited thereto. It should be noted that, in the process of starting the training of the object identification model 141, the computing unit 110 uses the simulated object image in the simulated object image collection. Since the simulated object image is obtained by simulating different scenes and different training objects (for example, analog license plate images), it can greatly cover the situation in the field to be tested and cannot be obtained, so the arithmetic unit 110 can not use the real image. Instead of the image of the scene, the simulated object image in the simulated object image set is used to train the object identification model 141.

In an embodiment, after the training of the object identification model 141 is completed, the computing unit 110 can capture the scene to be tested (for example, a scene with a vehicle) from an external host or captured by the image capturing device 120. The image to be tested is input into the object identification model 141 to obtain an object recognition result, wherein the object identification result is, for example, a license plate number in the image to be tested.

In another embodiment, the fifth database 139 of the non-volatile memory 132 stores a test image set including a plurality of test images, wherein the test image set may also be referred to as an unlabeled test image. set. The above test image is, for example, an image obtained by photographing a real scene including a vehicle and its license plate. For example, the computing unit 110 can input each test image in the test image set into the object identification model 141 to obtain a corresponding object identification result, and store the corresponding object identification result of each test image into the non-volatile memory 132. Five databases 139. Optionally, the computing unit 110 may mark the corresponding object identification result of each test image on each test image, and store the labeled test image separately into the sixth database 140 in the non-volatile memory 132.

In an embodiment, the identification result of the object identification model 141 is not 100% accurate due to various environmental changes, so the user can check the object identification of each test image in the test image set by manual inspection. The result is correct. If it is determined that the object identification result corresponding to a specific test image is incorrect, the operation unit 110 may add the specific test image to the fourth database 138, and input the correct object identification result corresponding to the specific test image to the object identification model. 141, by retraining and updating the analyte identification model 141, the recognition rate of the analyte identification model 141 in a similar situation can be improved. Similarly, if the object recognition result after the image to be tested captured by the scene to be tested is input to the object identification model 141 is incorrect, the operation unit 110 may add the image to be tested to the fourth database 138, and The correct object identification result corresponding to the specific test image is input to the object identification model 141, thereby retraining and updating the object identification model 141.

In another embodiment, the user may first store the test images and the correct object identification results in the fifth database 139. After the operation unit 110 trains the object identification model 141 in the initial stage, each test image in the fifth database 139 can be input into the object identification model 141 to generate an object identification result, and the generated object identification result is generated. Compare with the correct object identification results stored in advance. If the generated object identification result does not match the pre-stored correct object identification result (that is, the object recognition result is “failed”), the operation unit 110 may add the corresponding test image of the generated object identification result to the fourth database. 138, and input the correct object identification result into the object identification model 141, so as to retrain and update the object identification model 141, so that the recognition rate of the object identification model 141 can be improved.

In detail, the training process of the object identification model 141 in the present invention is mainly based on the simulated object image, and can use the image to be tested of the real scene or the test image in the fifth database 139 to assist in the correction and update. The analyte identification model 141.

In an embodiment, the image of the object captured in the real scene (such as the license plate image) may be visually different from the image of the environment, which is the object image feature described above, and may also be referred to as the test object. Objects (such as license plates) are special. Object image features include, for example, blurring, scratches or stains, shadows, shadowing, overexposure, distortion, and chromatic aberration. For example, different image representations can be used for each object image feature.

For example, in the case of a fuzzy feature, when the vehicle speed is too fast, the focus fails, or the vehicle is too far away, the license plate image may be blurred. Thus, the blurring feature can be represented, for example, by a blur mask, such as a matrix of M*N, and multiplied by a matrix of M*N at the center pixel of the blur mask to obtain a blurred center pixel. For example, the three columns of pixels of the license plate image in the blur mask are from left to right, from top to bottom, for example, a1 to a3, b1 to b3, and c1 to c3, respectively, where b2 is the center pixel, such as the 3A. The figure shows. The blur mask can be, for example, a 3x3 matrix, as shown in Figure 3B. The coefficient values in the 3x3 matrix are, for example, one, but the present invention is not limited to the above-described fuzzy mask, and the present invention can also be processed by the conventional fuzzy mask in the field of the present invention. Therefore, the central pixel b2 processed by the blur mask is updated to b2 = (a1*1 + a2*1 + a3*1 + b1*1 + b2*1 + b3*1 + c1*1 + c2*1 + C3*1)* (1/9).

For example, in the case of scratches or stains, the characters on the license plate may be scratched or stained, and the scratches may exist, for example, in a straight line or a curved line, and the stain may exist in a plane. Therefore, the arithmetic unit 110 can respectively use the straight line equation or the curve equation to simulate the scratch on the license plate, and simulate the stain on the license plate in a plane equation.

In the case of shadow features, the light source and the environment cause shadows in specific areas of the license plate image. Therefore, the arithmetic unit 110 can use a brightness mask for the license plate image to produce a shadow image effect. For example, if the three columns of the license plate image in the brightness mask are from top to bottom, for example, a1 to a3, b1 to b3, and c1 to c3, respectively, where b2 is the center pixel, as shown in FIG. 3A. The brightness mask can be, for example, a 3×3 matrix, as shown in FIG. 3C, wherein the three columns of the luminance mask are from left to right, from top to bottom, for example, h1 to h3, i1 to i3, and j1 to j3, respectively. The values of h1 to h3, i1 to i3, and j1 to j3 may be positive numbers greater than 1, or less than/equal to 1, depending on the design requirements of the brightness mask. Therefore, the arithmetic unit 110 can update the license plate image pixel a1 in the brightness mask to a1=a1*h1, the license plate image pixel a2 to a2=a2*h2, and so on.

For example, in the case of sheltering features, weather (dust, rain, snow) or other objects (leaves, insects, etc.) covering the license plate will produce a shadowing effect. Therefore, the arithmetic unit 110 can use one or more plane equations as a mask to shield a part of the area of the license plate image, and the size of the mask is not to destroy the characters on the license plate.

Taking the overexposure feature as an example, the light source from the lamp can not be suppressed and the area near the lamp is overexposed. Therefore, the arithmetic unit 110 can use a brightness mask for the license plate image to generate an overexposed image effect. For example, if the three columns of the license plate image in the brightness mask are from top to bottom, for example, a1 to a3, b1 to b3, and c1 to c3, respectively, where b2 is the center pixel, as shown in FIG. 3A. The brightness mask can be, for example, a 3×3 matrix, as shown in FIG. 3C, wherein the three columns of the luminance mask are from left to right, from top to bottom, for example, h1 to h3, i1 to i3, and j1 to j3, respectively. The values of the parameters h1 to h3, i1 to i3, and j1 to j3 may be positive numbers greater than 1, or less than/equal to 1, depending on the design requirements of the brightness mask, and parameters for the brightness mask of the overexposed feature. The value is different from the parameter value for the brightness mask of the shadow feature. Therefore, the arithmetic unit 110 can update the license plate image pixel a1 in the brightness mask to a1=a1*h1, the license plate image pixel a2 to a2=a2*h2, and so on.

Taking the deformation feature as an example, the different perspectives of the camera will produce three axes (X-axis, Y-axis, Z-axis) rotation of the captured license plate image. Therefore, the operation unit 110 can use a perspective transformation matrix for the license plate image to produce a deformed image effect. For example, the operation unit 110 can calculate a transparent transformation matrix according to Equation (1), for example: (1)

The operation unit 110 can set the values of the parameters a ₁₁ to a ₃₃ in the 3×3 matrix according to requirements, and pass the simulated object (for example, the simulated license plate composed of different characters) through the perspective transformation matrix (for example, (x, y) instead of the pixel value) of pixel values (x '/ w', y '/ w') of) the image plate to simulate different views.

Taking the color difference characteristic as an example, the photo opportunity is affected by the environment, and the light source is imaged by the lens to generate a color deviation. Therefore, the operation unit 110 can perform a color space conversion on the license plate image to achieve the image effect of the color difference.

In an embodiment, the background image features, for example, may be noise generated by images captured in a real scene, and the background image features may also be referred to as environmental noise features. Background image features include, for example, blurring, scratches or stains, shadows, noise, shadowing, overexposure, distortion, and chromatic aberration, but the invention is not limited thereto. For each background image feature, for example, different representations can be used. It should be noted that some of the image features are the same in the image image feature and the background image feature. The image features with the same name are treated similarly, but the image features are specific to each training object (for example, analog license plate image). For processing, the background image feature is processed for the entire background image (may not include the license plate) or the analog composite image, so the object image features and background image characteristics are different in setting the corresponding types of masks, matrices, and equations.

In one embodiment, the background image feature contains noise characteristics as compared to the image features of the object. For example, the computing unit 110 can add different types of noise to the image to be processed (for example, training objects, background images, or analog synthetic images), such as salt-and-pepper noise, Gaussian noise. ), speckle noise, or periodic noise. For the salt and pepper noise, the computing unit 110 can set the salt and pepper noise as x% of the image area of the image to be processed, and randomly add the salt and pepper noise to the image to be processed, wherein the value of x can be adjusted according to actual conditions. Regarding the Gaussian noise, the spot noise, and the periodic noise, the operation unit 110 can add these noises to the image to be processed by using a conventional technique, so the details thereof will not be described in detail herein.

4A-4F are schematic views showing training objects used in the object identification process in accordance with another embodiment of the present invention. In another embodiment, the training object generated by the computing unit 110 is not limited to the simulated license plate image. For example, training objects also include human bodies, license plates, components, and signs. In this embodiment, the image of the plurality of objects stored in the first database 135 is, for example, a human body image including one or more human postures, and the computing unit 110 identifies the human body region from the image of each object scene. The image is taken as an object image, and the captured object image is stored in the third database 137.

As shown in Figures 4A-4F, the image of the object in the third database 137 can be, for example, a human image obtained at different backgrounds and locations. In this embodiment, the predetermined rule is, for example, that the object image in the third database 137 can be directly used as the training object, so the operation unit 110 can directly select from the plurality of object images stored in the third database 137. One of them is used as a training object. In some embodiments, the predetermined rule may be, for example, arranging different one or more object images in a predetermined manner or pitch to produce a training article, but the invention is not limited thereto. Similarly, when the object to be recognized is a character, a component, or a logo, etc., the present invention can also store an image of the corresponding type of object in the first database 135, and extract the image of the object from the image of the object scene, and The simulation object image of the corresponding type is generated by using the process of the foregoing embodiment to form a simulated object image set, and the sample object recognition model 141 is trained according to the simulated object image set.

Figure 5 is a flow chart showing an object recognition method using a simulated object image in accordance with an embodiment of the present invention. Please also refer to Figures 1 and 5.

In step S510, an object image set including one of the plurality of object images and a background image set including one of the plurality of background images are obtained. The object image collection is stored, for example, in the third database 137, and the object image may be, for example, an image including one or more types of objects, such as text, human body, license plate, components, and logos, etc., but The invention is not limited to this. The background image collection is stored, for example, in the second database 136. The background image may be, for example, a real background image of any real scene obtained under different shooting conditions, and is not limited to the background image of the scene to be tested, and may not include the object to be tested. In some embodiments, the background image further includes a virtual background image simulated by computer vision technology.

In step S520, a set of simulated object images including a plurality of simulated object images is generated according to the object image set and the background image set. For example, the computing unit 110 composes one or more training objects from the one or more object images according to a predetermined rule, and performs a first image processing to add one or more object image features to the one or more trainings. Each of the objects is used to generate one or more simulated objects to be tested. The computing unit 110 can generate a simulated object image set according to one or more simulated objects to be tested and a background image set. The one or more object image features may be obtained, for example, from an image of the object scene in the first database 135, or may be simulated by an equation or matrix to simulate an image feature of the object of the training object. The computing unit 110 then obtains the first background image from the background image set in the second database 136, and performs a second image processing to add the one or more background images to the first background image to generate a simulated background image. The computing unit 110 can generate a simulated object image set according to, for example, one or more simulated objects to be tested and a simulated background image. Next, the computing unit 110 performs an image synthesis process to add the simulated object to be tested to the simulated background image to generate a simulated composite image, and performs the second image processing to add the one or more background images to the simulated composite image. To generate one of the images of the simulated objects.

In step S530, a simulated object recognition model is trained according to the simulated object image set. For example, in an embodiment, the computing unit 110 may first train the object identification model 141 through the simulated object image collection (ie, it is not necessary to use the real image for training). In another embodiment, the computing unit 110 can directly add the real object image to the simulated object image set to generate a mixture component image set, and train the DUT identification model according to the mixture component image set.

In step S540, a to-be-measured image obtained by a scene to be tested is input into the object recognition model to obtain an object recognition result. For example, the user may first store the test images and the correct object identification results in the fifth database 139. After the operation unit 110 trains the object identification model 141 in the initial stage, each test image in the fifth database 139 can be input into the object identification model 141 to generate an object identification result, and the generated object identification result is generated. Compare with the correct object identification results stored in advance. If the generated object identification result does not match the pre-stored correct object identification result (meaning the object recognition result is "failed"). In addition, when the object identification result of the object to be tested 141 is unsuccessful, the operation unit 110 may add the image to be tested to the simulated object image set to generate a mixture image image, and according to the mixture image The set and the correct object identification result of the image to be tested are used to train the test object identification model 141.

In summary, the present invention provides an object recognition system and method for simulating an object image, which can extract object features and environmental features using a small amount of data images, and thereby generate a large number of simulated object images and simulated backgrounds. Imagery that increases the diversity of training material collections (eg, simulated object image collections). Since the analog data is close to the actual data, the method can simulate the data mainly, supplemented by the real data, greatly reducing the time for data preparation and improving the dilemma encountered when the data is difficult to obtain.

The method of the present invention, or a specific type or part thereof, may be included in a physical medium such as a floppy disk, a compact disc, a hard disk, or any other machine (for example, a computer readable computer). A storage medium in which, when the code is loaded and executed by a machine, such as a computer, the machine becomes a device or system for participating in the present invention. The method, system and apparatus of the present invention may also be transmitted in a coded form via some transmission medium, such as a wire or cable, optical fiber, or any transmission type, wherein the code is received and loaded by a machine, such as a computer. And when executed, the machine becomes a device or system for participating in the present invention. When implemented in a general purpose processor, the code in conjunction with the processor provides a unique means of operation similar to application specific logic.

The present invention has been disclosed in the above preferred embodiments, and is not intended to limit the scope of the present invention. Any one of ordinary skill in the art can make a few changes without departing from the spirit and scope of the invention. The scope of protection of the present invention is therefore defined by the scope of the appended claims.

100‧‧‧ Object Identification System

110‧‧‧ arithmetic unit

120‧‧‧Image capture device

130‧‧‧storage unit

131‧‧‧ volatile memory

132‧‧‧ Non-volatile memory

133‧‧‧object identification program

135‧‧‧First database

136‧‧‧Second database

137‧‧‧ third database

138‧‧‧ Fourth database

139‧‧‧ Fifth database

140‧‧‧ Sixth Database

141‧‧‧Test object identification model

150‧‧‧ display

S510-S540‧‧‧Steps

1 is a block diagram showing an object recognition system in accordance with an embodiment of the present invention. The 2A-2M diagram shows a schematic diagram of different images used in the object recognition process in accordance with an embodiment of the present invention. 3A is a schematic diagram showing pixels of a training object in a blurred mask in accordance with an embodiment of the present invention. Figure 3B is a schematic diagram showing the coefficients in a blur mask in accordance with an embodiment of the present invention. Figure 3C is a schematic diagram showing the coefficients in a luminance mask in accordance with an embodiment of the present invention. 4A-4F are schematic views showing an image of an object according to another embodiment of the present invention. Figure 5 is a flow chart showing an object recognition method using a simulated object image in accordance with an embodiment of the present invention.

Claims (14)

An object recognition method using a simulated object image, the method comprising: (A) acquiring an image collection of an object including one or more object images and a background image set including one or more background images; (B) according to the object The image collection and the background image collection generate a simulation object image set including one of a plurality of simulated object images; (C) training a test object recognition model according to the simulated object image set; and (D) being subjected to a scene to be tested The obtained image to be tested is input into the object identification model to obtain an object identification result. The method for identifying an object using a simulated object image according to claim 1, wherein the step (B) comprises: forming one or more training objects from the one or more object images according to a predetermined rule; The first image processing adds one or more object image features to each of the one or more training objects to generate one or more simulated objects to be tested; and according to the one or more simulated objects to be tested and the A collection of background images to produce a collection of simulated object images. An object identification method using a simulated object image as described in claim 2, wherein the one or more object image features are obtained from the image of the object. The method for identifying an object using a simulated object image as described in claim 2, wherein the step (B) further comprises: obtaining a first background image from the one or more background images; performing a second image processing The one or more background images are specifically added to the first background image to generate an analog background image; and the simulated object image set is generated according to the simulated background image and the one or more simulated objects to be tested. The method for identifying an object using a simulated object image according to the fourth aspect of the patent application, wherein the step (B) further comprises: performing an image synthesis process to add the simulated object to be tested to the simulated background image to generate a simulated synthetic image. And performing the second image processing to add the one or more background images to the simulated composite image to generate one of the simulated object images. The method for identifying an object using a simulated object image as described in claim 1 further includes: (E) adding the image to be tested to the simulated object image set to generate a mixture when the object identification result is a failure. And acquiring (F) the training object identification model according to the image collection of the mixture and the correct object identification result of the image to be tested. The method for identifying an object using a simulated object image according to the first aspect of the invention, wherein the step (C) further comprises: adding one or more real object images to the simulated object image set to generate a mixture image collection; And training the test object identification model according to the image set of the mixture. An object recognition system using a simulated object image, comprising: a non-volatile memory for storing an object recognition program; and an arithmetic unit for executing the object recognition program to perform the following steps: (A) obtaining a complex number (1) generating a set of simulated object images including a plurality of simulated object images according to the object image set and the background image set; (C) And (D) inputting a to-be-measured image obtained by a scene to be tested into the object identification model to obtain an object identification result. An object identification system using a simulated object image according to claim 8 , wherein in the step (B), the computing unit further comprises one or more objects to form one or more training objects according to a predetermined rule. And performing a first image processing to add the one or more object image features to each of the one or more training objects to generate one or more simulated objects to be tested, and the computing unit is further based on the one or A plurality of simulated objects to be tested and the background image set to generate the simulated object image collection. An object recognition system using a simulated object image according to claim 9, wherein the one or more object image features are obtained from the image of the object. An object recognition system using a simulated object image according to claim 9, wherein in the step (B), the computing unit further acquires a first background image from the background images, and performs a second image processing. The one or more background image features are added to the first background image to generate a simulated background image, and the computing unit further generates the simulated object image set according to the simulated background image and the one or more simulated objects to be tested. An object recognition system using a simulated object image according to claim 11, wherein in the step (B), the operation unit further performs an image synthesis process to add the simulated object to be tested to the simulated background image to generate a Simulating the synthesized image and performing the second image processing to add the one or more background images to the simulated synthetic image to generate one of the simulated object images. The object recognition system using the simulated object image as described in claim 8 , wherein the operation unit further performs the following steps: (E) adding the image to be tested to the simulated object when the object identification result is a failure. The image is assembled to generate a mixture of image images; and (F) the object identification model is trained based on the image collection of the mixture and the correct object identification result of the image to be tested. An object recognition system using a simulated object image according to claim 8 , wherein in the step (C), the computing unit further adds one or more real object images to the simulated object image set to generate a mixture piece. The image collection is based on the mixture image collection to train the object identification model.