System and method for generating multiple-exposure frames from single input

Abstract

A method includes obtaining, using at least one image sensor of an electronic device, a first image frame of a scene. The method also includes generating, using a convolutional neural network, a plurality of second image frames simulated to have different exposures from the first image frame. One or more objects in the scene in each second image frame are aligned with one or more corresponding objects in the scene in at least one other second image frame and aligned with one or more corresponding objects in the scene in the first image frame. The method also includes blending the plurality of second image frames to generate a final image of the scene.

Description

TECHNICAL FIELD

[0001] The present disclosure relates generally to image capture systems. More specifically, the present disclosure relates to a system and method for generating multiple exposure frames from a single input. BACKGROUND

[0002] Many mobile electronic devices, such as smartphones and tablets, include cameras that can be used to capture still and video images. Despite the convenience, cameras on mobile electronic devices often suffer from a number of shortcomings. For example, cameras on mobile electronic devices often capture images with underexposed or overexposed regions, such as when capturing images of natural scenes. This is often because the dynamic range of the image sensor in the camera is limited. Multiple image frames of a scene can be captured and then the“best” portions of the image frames combined to generate a blended image. However, generating a blended image from a set of image frames with different exposures is a challenging process, particularly for dynamic scenes. As another example, cameras on mobile electronic devices often perform poorly in low light conditions. While the amount of light collected at the image sensor can be increased by increasing the exposure time, this also increases the risk of producing a blurred image due to motion of the objects and the camera. SUMMARY

[0003] The present disclosure provides a system and method for generating multiple exposure frames from a single input.

[0004] In a first embodiment, a method includes obtaining, using at least one image sensor of an electronic device, a first image frame of a scene. The method also includes generating, using a convolutional neural network, a plurality of second image frames simulated to have different exposures from the first image frame. One or more objects in the scene in each second image frame are aligned with one or more corresponding objects in the scene in at least one other second image frame and aligned with one or more corresponding objects in the scene in the first image frame. The method further includes blending the plurality of second image frames to generate a final image of the scene.

[0005] In a second embodiment, an electronic device includes at least one image sensor and at least one processing device. The at least one processing device is configured to obtain, using the at least one image sensor, a first image frame of a scene. The at least one processing device is also configured to generate, using a convolutional neural network, a plurality of second image frames simulated to have different exposures from the first image frame. One or more objects in the scene in each second image frame are aligned with one or more corresponding objects in the scene in at least one other second image frame and aligned with one or more corresponding objects in the scene in the first image frame. The at least one processing device is further configured to blend the plurality of second image frames to generate a final image of the scene.

[0006] In a third embodiment, a non-transitory machine-readable medium includes instructions that, when executed, cause at least one processor of an electronic device to obtain, using at least one image sensor of the electronic device, a first image frame of a scene. The medium also includes instructions that, when executed, cause the at least one processor to generate, using a convolutional neural network, a plurality of second image frames simulated to have different exposures from the first image frame from the first image frame. One or more objects in the scene in each second image frame are aligned with one or more corresponding objects in the scene in at least one other second image frame and aligned with one or more corresponding objects in the scene in the first image frame. The medium also includes instructions that, when executed, cause the at least one processor to blend the plurality of second image frames to generate a final image of the scene.

[0007] Other technical features can be readily apparent to one skilled in the art from the following figures, descriptions, and claims.

[0008] Before undertaking the detailed description below, it can be advantageous to set forth definitions of certain words and phrases used throughout this patent document. The terms “transmit,” “receive,” and “communicate,” and variations thereof, encompass both direct and indirect communication. The terms “include” and “comprise,” as well as variations thereof, mean “including but not limited to.” The term “or” is inclusive, meaning and / or. The phrase “associated with,” as well as variations thereof, means in connection with; included within; in communication with; interconnected with; contains; comprised within; has a property of; possesses; or any kind of input or output between any two entities.

[0009] Also, various functions described below can be implemented or supported by one or more computer programs, each of which is formed from computer readable program code and embodied in a computer readable medium. The terms “application” and “program” refer to one or more computer programs, software components, sets of instructions, procedures, functions, objects, classes, instances, related data, or a portion thereof. The phrase “computer readable program code” includes any type of computer code, including source code, object code, and executable code. The phrase “computer readable medium” includes any type of medium capable of being accessed by a computer, such as read only memory (ROM), random access memory (RAM), a hard disk drive, a compact disc (CD), a digital video disc (DVD), or any other type of memory. A “non-transitory” computer readable medium excludes wired, wireless, optical, or other communication links. Non-transitory computer readable media include media where data is permanently stored and media where data is stored and later overwritten, such as a rewritable optical disc or an erasable memory device.

[0010] As used herein, terms and phrases such as “have,” “has,” “can,” “having,” “include,” “including,” “comprise,” “comprising,” “containing,” “contain” or variations of these or similar terms, shall not be a limitation on the terms of this disclosure but shall be read to include other situations if these situations are within the scope of the equivalent terminology. Also, as used herein, the phrases “A or B,” “at least one of A and / or B,” or “one or more of A and / or B” can include all possible combinations of A and B. For example, “A or B,” “at least one of A and B,” and “at least one of A or B” can indicate (1) at least one A, (2) at least one B, or (3) at least one A and at least one B. Further, as used herein, the terms “first,” “second,” and the like, can modify various components, regardless of importance and cannot exclude the components from the scope of the disclosure. The terms can be used to distinguish a component from another component with a similar or same name or to distinguish a component from another component. For example, a first user device and a second user device can indicate different user devices from each other regardless of order or importance of the devices. The first component can be denoted as the second component, and vice versa, without departing from the scope of the disclosure.

[0011] It will be understood that when an element (for example, a first element) is referred to as being “coupled with / to” or “connected with / to” another element (for example, a second element) or “connected with / to” another element, it can be directly coupled or connected with / to the other element or coupled or connected with / to the other element via a third element. In contrast, it will be understood that when an element (for example, a first element) is referred to as being “directly coupled with / to” or “directly connected with / to” another element (for example, a second element), no other element (for example, a third element) is interposed therebetween.

[0012] As used herein, the phrase “configured (or set) to” can be used interchangeably with phrases “adapted to,” “capable of,” “designed to,” “suitable for,” “made to,” or “to be able to,” depending on the context. The phrase “configured (or set) to” does not literally mean “designed in hardware to” in essence. Rather, the phrase “configured to” can mean that a device can perform an operation with another device or component. For example, the phrase “a processor configured (or set) to perform A, B, and C” can mean a general-purpose processor (for example, a CPU or an application processor) that can perform the operations by executing one or more software programs stored in a memory device or a dedicated processor (for example, an embedded processor) for performing the operations.

[0013] The terminology and phraseology used herein is solely used for the purpose of describing some embodiments of the present disclosure and is not intended to limit the scope of other embodiments of the present disclosure. It is to be understood that the singular forms "a," "one," and "the" include plural referents unless the context clearly dictates otherwise. All terminology and phrases used herein (including technical and scientific terminology and phraseology) have the same meaning as commonly understood by one of ordinary skill in the art to which embodiments of the present disclosure belong. It will also be understood that terms and phrases such as those defined in commonly used dictionaries should be construed in a manner that is consistent with their meaning in the context of the relevant art and will not be interpreted in an idealized or overly formal sense unless expressly so defined herein. In some instances, terms and phrases defined herein can be interpreted as excluding embodiments of the present disclosure.

[0014] Examples of the "electronic device" according to an embodiment of the present disclosure can include at least one of a smartphone, a tablet personal computer (PC), a mobile phone, a video phone, an e-book reader, a desktop PC, a laptop computer, a netbook computer, a workstation, a personal digital assistant (PDA), a portable multimedia player (PMP), an MP3 player, a mobile medical device, a camera, or a wearable device (e.g., smart glasses, a head-mounted device (HMD), electronic clothes, an electronic bracelet, an electronic necklace, an electronic appcessory, an electronic tattoo, a smart mirror, or a smart watch). Other examples of the electronic device include a smart home appliance. Examples of the smart home appliance can include at least one of a television, a digital video disc (DVD) player, an audio player, a refrigerator, an air conditioner, a cleaner, an oven, a microwave oven, a washing machine, a dryer, an air cleaner, a set-top box, a home automation control panel, a security control panel, a TV box (e.g., SAMSUNG HOMESYNC, APPLE TV, or GOOGLE TV), a smart speaker or a speaker with an integrated digital assistant (e.g., SAMSUNG GALAXY HOME, APPLE HOMEPOD, or AMAZON ECHO), a game console (e.g., XBOX, PLAYSTATION, or NINTENDO), an electronic dictionary, an electronic key, a camcorder, or an electronic frame. Further examples of the electronic device include at least one of various medical devices (e.g., various portable medical measuring devices (e.g., a blood glucose measuring device, a heart rate measuring device, or a body temperature measuring device), a magnetic resonance angiography (MRA) device, a magnetic resonance imaging (MRI) device, a computed tomography (CT) device, an imaging device, or an ultrasonic device), a navigation device, a global positioning system (GPS) receiver, an event data recorder (EDR), a flight data recorder (FDR), a car infotainment device, a sailing electronic device (e.g., a sailing navigation device or a gyro compass), avionics, security devices, a car head unit, an industrial or household robot, an automatic teller machine (ATM), a point of sale (POS) device, or an Internet of Things (IoT) device (e.g., a light bulb, various sensors, a gas or water meter, a sprinkler, a fire alarm, a thermostat, a street light, a toaster, a fitness device, a hot water tank, a heater, or a boiler). Further examples of the electronic device include at least one of furniture or a portion of a building / structure, an electronic board, an electronic signature receiving device, a projector, or various measuring devices (e.g., a device for measuring water, electricity, gas, or electromagnetic waves). Note that the electronic device according to various embodiments of the present disclosure can be one or a combination of the above-listed devices. According to some embodiments of the present disclosure, the electronic device can be a flexible electronic device. The electronic device disclosed herein is not limited to the above-listed devices and can include new electronic devices according to technological development.

[0015] In the following description, an electronic device according to various embodiments of the disclosure is described with reference to the accompanying drawings. As used herein, the term "user" can represent a person or another device (e.g., an artificial intelligence electronic device) using the electronic device.

[0016] Definitions for other certain words and phrases are provided throughout this patent document. Those of ordinary skill in the art will understand that in many, if not most instances, such definitions apply to not only the

[0017] No description in the present application should be interpreted as implying any particular element, step or function is an essential element that must be included in the claim scope. The scope of the patent subject matter is defined only by the claims. Moreover, unless the exact phrase "means for" followed by a statement of function is recited, no claim element is intended to be invoked as a means-plus-function clause. As used herein, the term "and / or" includes any and all combinations of one or more of the associated listed items. In addition, unless the exact phrase "device for" followed by a statement of function is recited, no claim element is intended to be invoked as a means-plus-function clause. With respect to the use of any other term or phrase as a means-plus-function clause, such clause is intended to cover the equivalent structures of what is described irrespective of structure or terminology. The term "about" when used before a statement regarding an amount or a numerical value is intended to refer to plus or minus 10% of the stated value. BRIEF DESCRIPTION OF DRAWINGS

[0018] For a more complete understanding of the present disclosure and its advantages, reference is now made to the following description taken in conjunction with the accompanying drawings in which like reference numerals represent like parts:

[0019] Figure 1 An example network configuration including an electronic device according to this disclosure is illustrated;

[0020] Figure 2 An example process for generating multiple image frames using a single input multiple exposure (SIME) network according to this disclosure is illustrated;

[0021] Figure 3 Additional details of a SIME network according to this disclosure are illustrated; Figure 2

[0022] Figure 4A And Figure 4B Example options for implementing a SIME network according to this disclosure are illustrated; Figure 3

[0023] Figure 5 An example process for training a SIME network according to this disclosure is illustrated;

[0024] Figure 6 ​​An example process for generating artifact-free HDR frames using multiple output image frames generated by a SIME network according to the present disclosure is shown;

[0025] Figures 7A-7D Another example process for generating artifact-free HDR frames using multiple output image frames generated by a SIME network according to the present disclosure is shown;

[0026] Figure 8 Yet another example process for generating artifact-free HDR frames using multiple output image frames generated by a SIME network according to the present disclosure is shown;

[0027] Figure 9 An example process for changing an image domain using a SIME network according to the present disclosure is shown;

[0028] Figure 10 Another example process for changing an image domain using a SIME network according to the present disclosure is shown;

[0029] Figure 11A , Figure 11B , Figure 12A and Figure 12B An example of the benefits that can be realized using one or more embodiments of the present disclosure is shown; and

[0030] Figure 13 An example method for generating multiple exposure frames using a SIME network according to the present disclosure is shown. DETAILED DESCRIPTION

[0031] The following discussion is presented to enable a thorough and complete understanding of the various embodiments of the present disclosure, as well as the various features, functionalities, and concepts thereof. However, it will be understood that the Figures 1-13 and various embodiments of the present disclosure. However, it is to be understood that the present disclosure is not limited to these embodiments, and that all changes and / or equivalents or alternatives coming within the spirit and scope of the present disclosure are encompassed by the present disclosure. Throughout the specification and drawings, like or similar elements can be referred to with the same or similar reference designators.

[0032] As noted above, many mobile electronic devices, such as smartphones and tablets, include cameras that can be used to capture still and video images, but these cameras suffer from a number of shortcomings. For example, such as when capturing images of natural scenes, these cameras often capture images having underexposed or overexposed regions. As another example, these cameras typically perform poorly in low light conditions, and increasing the exposure time also increases the risk of motion blur.

[0033] Many imaging applications require the use of multiple image frames captured at different exposure levels. Example types of imaging applications herein can include applications that support high dynamic range (HDR), relighting, and the like. Traditionally, multiple image frames captured at different exposure levels are captured sequentially, and thus at different times. As a result, one or more objects in the scene can move between image frames, for example due to motion of the objects or motion of the camera. This can cause difficulties in the imaging application, particularly where perfect (or near perfect) alignment between the image frames is preferred or required.

[0034] In many multi-frame applications, misalignment between image frames presents a significant challenge to multi-frame processing and can result in the creation of various types of image artifacts (e.g., blended ghosting, attenuated blending, speckle noise, etc.). As a particular example, blended images can contain one or more blurred portions in areas where motion occurred. While some artifacts (e.g., ghosting) can be somewhat hidden by over-brightening surrounding areas, this can be undesirable, for example when the over-brightening hides details of background objects and reduces the overall sharpness of the image.

[0035] The present disclosure provides various techniques for generating multiple-exposure image frames from a single input. As described in greater detail below, embodiments of the present disclosure provide neural network-based architectures that generate multiple image frames at different exposure levels based on a single captured image frame. The disclosed neural network-based architectures (also referred to below as single-input multiple-exposure or “SIME” networks) can efficiently generate multiple-exposure image frames based on a single input image frame. The generated multiple-exposure image frames are aligned in multiple image domains, such as the Bayer (“raw format”) domain and the YUV (“visual format”) domain. The multiple-exposure image frames can then be blended or otherwise processed to generate one or more images of the scene with improved image quality, for example with fewer or no under / over-exposed regions and little or no motion blur. These techniques provide visually better image frames compared to traditional image signal processor-based approaches, for example those that apply gain and noise filtering to captured image frames.

[0036] Note that while the techniques described below are often described as being performed using a mobile electronic device, other electronic devices can also be used to perform or support these techniques. Thus, these techniques can be used in various types of electronic devices. Also, while the techniques described below are often described as processing image frames when capturing still images of a scene, the same or similar approaches can be used to support the capture of video images.

[0037] Figure 1 An example network configuration 100 including an electronic device according to the present disclosure is shown. Figure 1The illustrated embodiment of the network configuration 100 is for illustration only. Other embodiments of the network configuration 100 can be used without departing from the scope of the present disclosure.

[0038] According to an embodiment of the present disclosure, the electronic device 101 is included in the network configuration 100. The electronic device 101 can include at least one of a bus 110, a processor 120, a memory 130, an input / output (I / O) interface 150, a display 160, a communication interface 170, or a sensor 180. In some embodiments, the electronic device 101 can exclude at least one of the components, or can add at least one other component. The bus 110 includes a circuit for connecting the components 120-180 to each other and for transmitting communications (e.g., control messages and / or data) between the components.

[0039] The processor 120 includes one or more of a central processing unit (CPU), an application processor (AP), or a communication processor (CP). The processor 120 is capable of performing control over at least one of the other components of the electronic device 101 and / or performing operations or data processing related to communication. In some embodiments, the processor 120 can be a graphic processing unit (GPU). For example, the processor 120 can receive image data captured by at least one camera during a capture event. Further, the processor 120 can process the image data using a convolutional neural network (as discussed in more detail below) to generate a plurality of aligned image frames.

[0040] The memory 130 can include a volatile and / or nonvolatile memory. For example, the memory 130 can store commands or data related to at least one of the other components of the electronic device 101. According to an embodiment of the present disclosure, the memory 130 can store software and / or a program 140. The program 140, for example, includes a kernel 141, middleware 143, an application programming interface (API) 145, and / or an application program (or "application") 147. At least a portion of the kernel 141, the middleware 143, or the API 145 can be denoted as an operating system (OS).

[0041] The kernel 141 can control or manage system resources (e.g., the bus 110, the processor 120, or the memory 130) used to execute operations or functions implemented in other programs (e.g., the middleware 143, the API 145, or the application programs 147). The kernel 141 provides an interface that allows the middleware 143, the API 145, or the application 147 to access the individual components of the electronic device 101 to control or manage the system resources. The application 147 includes one or more applications for image capture and image processing as discussed below. These functions can be performed by a single application or by multiple applications, each of which performs one or more of these functions. The middleware 143 can serve, for example, as a relay to allow the API 145 or the application 147 to communicate data with the kernel 141. A plurality of applications 147 can be provided. The middleware 143 is capable of, for example, controlling work requests received from the application 147 by assigning priorities for use of the system resources (e.g., the bus 110, the processor 120, or the memory 130) of the electronic device 101 to at least one of the plurality of applications 147. The API 145 is an interface that allows the application 147 to control functions provided from the kernel 141 or the middleware 143. The API 145 includes, for example, at least one interface or function (e.g., a command) for file control, window control, image processing, or text control.

[0042] The I / O interface 150 serves as, for example, an interface that can transfer a command or data input from a user or other external device to other components of the electronic device 101. The I / O interface 150 can also output a command or data received from other components of the electronic device 101 to the user or other external device.

[0043] The display 160 includes, for example, a liquid crystal display (LCD), a light emitting diode (LED) display, an organic light emitting diode (OLED) display, a quantum dot light emitting diode (QLED) display, a microelectromechanical system (MEMS) display, or an electronic paper display. The display 160 can also be a depth perception display, such as a multi-focal display. The display 160 is capable of, for example, displaying various contents (e.g., text, images, videos, icons, or symbols) to the user. The display 160 can include a touch screen and can receive, for example, touch, gesture, proximity, or hovering inputs using an electronic pen or a user's body part.

[0044] The communication interface 170 is capable of, for example, establishing communication between the electronic device 101 and an external electronic device (e.g., the first electronic device 102, the second electronic device 104, or the server 106). For example, the communication interface 170 can be connected with the network 162 or 164 through wireless communication or wired communication to communicate with the external electronic device. The communication interface 170 can be a wired or wireless transceiver or any other component for transmitting and receiving signals such as images.

[0045] The electronic device 101 also includes one or more sensors 180 that can measure physical quantities or detect an activation state of the electronic device 101 and convert the measured or detected information into an electrical signal. For example, the one or more sensors 180 include one or more cameras or other image sensors for capturing images of a scene. The sensors 180 can also include one or more buttons for touch input, gesture sensors, gyroscopes or gyro sensors, barometric sensors, magnetic sensors or magnetometers, acceleration sensors or accelerometers, grip sensors, proximity sensors, color sensors (e.g., red green blue (RGB) sensors), biophysical sensors, temperature sensors, humidity sensors, illuminance sensors, ultraviolet (UV) sensors, electromyography (EMG) sensors, electroencephalogram (EEG) sensors, electrocardiogram (ECG) sensors, infrared (IR) sensors, ultrasonic sensors, iris sensors, or fingerprint sensors. The sensors 180 can also include an inertial measurement unit, which can include one or more accelerometers, gyroscopes, and other components. In addition, the sensors 180 can include a control circuit for controlling at least one of the sensors included herein. Any of these sensors 180 can be located within the electronic device 101. The one or more cameras or other image sensors can optionally be used in conjunction with at least one flash 190. The flash 190 represents a device configured to generate illumination for use in image capture by the electronic device 101, such as one or more LEDs.

[0046] The first external electronic device 102 or the second external electronic device 104 can be a wearable device or a wearable device (e.g., an HMD) of a mountable electronic device. When the electronic device 101 is mounted in the electronic device 102 (e.g., an HMD), the electronic device 101 can communicate with the electronic device 102 through the communication interface 170. The electronic device 101 can be directly connected with the electronic device 102 to communicate with the electronic device 102 without involving a separate network. The electronic device 101 can also be an augmented reality wearable device, such as glasses, including one or more cameras.

[0047] The wireless communication can use at least one of, for example, Long Term Evolution (LTE), Long Term Evolution-Advanced (LTE-A), a 5th generation wireless system (5G), millimeter wave or 60 GHz wireless communication, wireless USB, Code Division Multiple Access (CDMA), Wideband Code Division Multiple Access (WCDMA), a Universal Mobile Telecommunication System (UMTS), Wireless Broadband (WiBro), or a Global System for Mobile Communications (GSM) as a cellular communication protocol. The wired connection can include, for example, at least one of a Universal Serial Bus (USB), a High Definition Multimedia Interface (HDMI), Recommended Standard 232 (RS-232), or a Plain Old Telephone Service (POTS). The network 162 or 164 includes at least one communication network, for example, a computer network (e.g., a Local Area Network (LAN) or a Wide Area Network (WAN)), the Internet, or a telephone network.

[0048] The first and second external electronic devices 102 and 104 and the server 106 can each be a device of the same or different type as the electronic device 101. According to certain embodiments of the present disclosure, the server 106 includes a group of one or more servers. Also, according to certain embodiments of the present disclosure, all or some of the operations performed on the electronic device 101 can be performed on another electronic device or a plurality of other electronic devices (e.g., the electronic devices 102 and 104 or the server 106). Furthermore, according to certain embodiments of the present disclosure, when the electronic device 101 is supposed to automatically or on request perform a certain function or service, instead of performing the function or service by itself or in addition to that, the electronic device 101 can request another device (e.g., the electronic devices 102 and 104 or the server 106) to perform at least some functions associated therewith. The other electronic device (e.g., the electronic devices 102 and 104 or the server 106) can perform the requested function or additional function and transmit the execution result to the electronic device 101. The electronic device 101 can provide the requested function or service by processing the received result as is or additionally. To this end, cloud computing, distributed computing, or client-server computing technology can be used, for example. Although Figure 1 It is illustrated that the electronic device 101 includes the communication interface 170 to communicate with the external electronic device 104 or the server 106 via the network 162 or 164, but according to some embodiments of the present disclosure, the electronic device 101 can operate independently without a separate communication function.

[0049] The server 106 can optionally support the electronic device 101 by performing or supporting at least one of the operations (or functions) implemented on the electronic device 101. For example, the server 106 can include a processing module or a processor that can support the processor 120 implemented in the electronic device 101.

[0050] Although Figure 1An example of a network configuration 100 including an electronic device 101 is shown, but other configurations are possible. Figure 1 Various changes can be made. For example, network configuration 100 can include any number of individual components arranged in any suitable manner. Typically, computing and communication systems have a wide variety of configurations, while... Figure 1 This disclosure is not intended to limit the scope to any particular configuration. Furthermore, although... Figure 1 An operating environment in which the various features disclosed in this patent document can be used is shown, but these features can also be used in any other suitable system.

[0051] Figure 2 An example process 200 for generating multiple image frames using a SIME network according to this disclosure is shown. For ease of illustration, Figure 2 The process 200 shown is described as involving the use of Figure 1 Electronic device 101. However, Figure 2 The process 200 shown can be used with any other suitable electronic device and can be used in any suitable system.

[0052] like Figure 2 As shown, electronic device 101 receives or acquires a single input image frame 201 of a scene captured using at least one camera or other image sensor of electronic device 101. In some embodiments, the input image frame 201 of the scene is captured using a short exposure time, at least relative to an automatically determined exposure time (often referred to as the EV-0 exposure time). Because of the short exposure time, the input image frame 201 tends to be dark, and not much detail is clearly visible in the input image frame 201.

[0053] An input image frame 201 (also referred to as a "short frame") is provided as input to a SIME network 205, which represents a convolutional neural network. An electronic device 101 performs or otherwise uses the SIME network 205 to generate multiple output image frames 211-213 using a single input image frame 201. Further details of the operation of the SIME network 205 are provided below. In this example, three output image frames 211-213 are generated using the SIME network 205. However, in other embodiments, the SIME network 205 may be used to generate two or more output image frames.

[0054] The output image frames 211-213 are generated to simulate images of the same scene captured at different exposure levels. In some embodiments, the first output image frame 211 can represent a short-exposure sharp image frame (also referred to as a "short sharp frame"). The output image frame 211 can be substantially similar to the input image frame 201, but most or all of the noise contained in the input image frame 201 can be removed by the SIME network 205 using a "denoising" process. The second output image frame 212 can represent a normal-exposure image frame (also referred to as a "normal frame"). The output image frame 212 can be simulated as having been captured using an automatically determined exposure time, and can be generated by the SIME network 205 applying a denoising process and a brightening process to the input image frame 201. The third output image frame 213 can represent a long-exposure image frame (also referred to as a "long frame"). The output image frame 213 can be simulated as having been captured using an exposure time that is longer than the automatically determined exposure time, and can be generated by the SIME network 205 applying both a denoising process and an even stronger brightening process to the input image frame 201. Note that the terms "short," "normal," and "long" as used here are relative to each other, and can represent any suitable exposure times such that "normal" is longer than "short," and "long" is longer than "normal." In many cases, "normal" refers to an automatically determined exposure time that results in an image having the least underexposed and / or overexposed areas.

[0055] Because multiple output image frames 211-213 are generated from a single input image frame 201, the output image frames 211-213 can exhibit better interframe alignment than traditional multi-frame methods. To a viewer, the output frames 211-213 appear to have perfect (or near perfect) alignment, without any blurring or movement. Additionally, the computational efficiency of using the SIME network 205 is higher compared to performing multiple single-input single-exposure (SISE) networks, which can be slower or require more processing power to run than the SIME network 205.

[0056] Although Figure 2 One example of a process 200 for generating multiple image frames using a SIME network is shown, but various changes can be made Figure 2 For example, while the process 200 is shown as generating three output image frames, other embodiments can produce more or fewer than three output image frames. Also, the operation of the SIME network 205 can be performed by any suitable component of the electronic device 101 or other device, including the processor 120 of the electronic device 101 or the image sensor 180 of the electronic device 101. In some cases, the SIME network 205 can be embedded in the image sensor itself.

[0057] Figure 3The following is shown in accordance with this disclosure: Figure 2 Additional details for SIME network 205. (e.g.) Figure 3 As shown, the SIME network 205 operates overall to receive the input image frame 201 and generate multiple output image frames 211-213. To achieve this, the SIME network 205 is based on a convolutional neural network architecture. A convolutional neural network architecture typically represents a deep artificial neural network commonly used for image analysis. In this example, the SIME network 205 includes an encoder path 302 and multiple decoder paths 304.

[0058] In this example, Figure 3 The numbers included in some components (e.g., 2, 4, 32, 64, 128, etc.) represent the number of channels, features, or components representing the input image frame 201. The encoding process typically increases the number of components representing the input image frame 201, meaning that the input image frame 201 is downsampled and more components are represented (but each component becomes spatially smaller). The decoding process typically decreases the number of components representing the output image frames 211-213, meaning that the output of the encoding process is upsampled and fewer components are represented (but each component becomes spatially larger). Figure 3 In this context, the symbol "+" is used to indicate the cascading of information (e.g., 256+128, 64+32, etc.). It should be noted that although in... Figure 3 Specific values ​​are shown, but these values ​​are for illustrative purposes only.

[0059] Encoder path 302 consists of multiple coding levels 306a-306d. Each of coding levels 306a-306c includes two convolutional layers 311 and a downsampling layer 312, and coding level 306d includes two convolutional layers 311 but no downsampling layer. In some embodiments, the two convolutional layers 311 may be generalized as a residual subnet, which may include skip connections on the two convolutional layers 311 and then perform an addition operation. Each convolutional layer 311 represents a layer of convolutional neurons operable to apply convolutional operations that simulate the response of a single neuron to a visual stimulus. Typically, each neuron applies a function to its input values ​​(typically by weighting different input values ​​differently) to generate an output value. Each downsampling layer 312 represents a layer that combines the output values ​​from a cluster of neurons in one layer into the input values ​​of the next layer. The encoder path 302 here is shown as including four coding levels 306a-306d with a total of eight convolutional layers 311 and three downsampling layers 312. However, encoder path 302 may include different numbers of coding levels, convolutional layers, and downsampling layers.

[0060] In some embodiments, each convolutional layer 311 can perform a convolution with a filter bank (containing multiple filters or kernels) to generate a set of feature maps. These feature maps can be batch-normalized, and an element-wise rectified linear unit (ReLU) function can be applied to the normalized feature map values. The ReLU function can generally operate to ensure that all output values of the layer are non-negative, e.g., by (for each normalized feature map value) selecting the greater of that value or zero. Subsequently, in some embodiments, each down-sampling layer 312 can perform max-pooling, e.g., with a window and a stride of 2 (non-overlapping windows), and the resulting output can be sub-sampled (e.g., by a factor of 2). Max-pooling can be used to achieve translational invariance over small spatial displacements in the input image frames 201, and sub-sampling can yield large input image contexts (spatial windows) for each pixel in the feature maps.

[0061] The common encoder path 302 is followed by multiple different decoder paths 304 for the multiple output image frames 211-213. Each decoder path 304 is formed using one or more decoder stages 308a-308c, where the number of decoder stages 308a-308c in each decoder path 304 depends on the type of output image frame 211-213 generated by that decoder path 304. For example, one decoder stage 308c can be used to generate the short clear output image frame 211, two decoder stages 308b-308c can be used to generate the normal output image frame 212, and three decoder stages 308a-308c can be used to generate the long output image frame 213.

[0062] Each decoder stage 308a-308c includes an up-sampling layer 313, which represents a layer that resizes its input. This can be achieved by upsampling the feature maps received from the previous stage and concatenating with the feature maps from the corresponding encoder stage. In Figure 3 The concatenation is depicted by the arrows connecting the up-sampling layer 313 with the previous stage and the corresponding encoder stage. Each decoder stage 308a-308c also includes a separable convolutional layer 314, which is used instead of an ordinary convolutional layer to efficiently reduce the complexity in the most computationally demanding part of the SIME network 205 (as it is compressed from a much higher depth). Each decoder stage 308a-308c also includes an ordinary convolutional layer 311 with moderate complexity, as the previous layer has the same depth. Since fine details are removed along the encoding path in a way that favors high-level features, a skip connection from an encoding layer located at the same scale to a decoding layer can be used to bring details back into the decoder paths 304. Figure 3

[0063] ​In this example, each decoder stage 308a-308c is executed to avoid the checkerboard artifacts typically observed in traditional transposed convolutional layers. That is, instead of using transposed convolutions, image resizing is performed first in each decoder stage 308a-308c, followed by the convolutional layer.

[0064] Depending on the implementation, various advantages can be achieved using the SIME network 205 based on the following principles and observations. Typically, the decoding path runs in cascades of previous layers, making decoding convolutional layers computationally expensive. A better trade-off between computation and quality can be achieved by replacing standard convolutional layers with separable convolutional layers 314. Furthermore, the SIME network 205 is more efficient because the three output image frames 211-213 share the encoder path 302, rather than requiring a separate encoder path 302 for each output image frame 211-213. Additionally, the decoder path 304 for the short, sharp output image frame 211 is smaller than the decoder path 304 for the long output image frame 213, because the smaller exposure gaps (brightness variations) require less intensive computation.

[0065] Furthermore, due to the large exposure gap between the short input image frame 201 and the long output image frame 213, a larger receiving field or deeper layers are used to optimize the trade-off between advanced noise suppression and signal preservation. As the exposure gap between input and output decreases, shorter exposure gaps require smaller receiving fields or higher-level structures to maintain the same level of trade-off between detail and noise suppression. In other words, smaller receiving fields and shallower layers are sufficient to achieve image quality optimization. By utilizing this observation, the SIME network 205 can generate multi-exposure image frames in a more computationally efficient manner compared to executing the SIME network multiple times.

[0066] although Figure 3 Some additional details of the SIME network 205 are shown, but more details can be found on the network itself. Figure 3 Various changes can be made. For example, the number of stages, convolutional layers, upsampling layers, downsampling layers, and other components can vary, and their number can be less or more.

[0067] Figure 4A and Figure 4B The present disclosure illustrates the method for performing Figure 3 Example options for SIME network 205. For example... Figure 4A and Figure 4B As shown, the SIME network 205 can operate according to different execution options 401-402 to generate output image frames in different formats. In other words, depending on the operations performed by the SIME network 205, different formats can exist for the output image frames 211-213.

[0068] In both options 401-402, the SIME network 205 receives an input image frame 201 in its raw format, containing minimally processed data from the image sensor. The raw format can also be associated with a Bayer image domain. As indicated by option 401, the SIME network 205 is operable to generate output image frames 211-213 in the same raw format as the input image frame 201. Conversely, as indicated by option 402, when the SIME network 205 generates output image frames 211-213, it performs a transformation operation to generate output image frames 211-213 in YUV format, a format commonly used for image storage. Training can be performed on the SIME network 205 to allow it to generate output images in either raw or YUV format.

[0069] although Figure 4A and Figure 4B It shows the method for execution Figure 3 Examples of options for SIME network 205, but more are available for... Figure 4A and Figure 4B Various modifications can be made. For example, the SIME network 205 can be compatible with other image formats such as RGB.

[0070] Training convolutional neural networks typically requires a large number of training examples. In the case of SIME network 205, the goal of training is to tune SIME network 205 so that when a single input image frame is provided to it, SIME network 205 generates multiple output image frames at different exposure levels.

[0071] Figure 5 An example procedure 500 for training a SIME network 205 according to this disclosure is shown. Using procedure 500, the SIME network 205 can be trained to generate multi-exposure frames that match a predetermined target frame based on only a single input frame. Constraints from the static scene and the single input frame ensure that the output frames generated by the SIME network 205 are perfectly aligned even at different exposure levels. For ease of illustration, procedure 500 is described as involving using Figure 1 Server 106 and Figure 2 and Figure 3 The SIME network 205 is shown. However, process 500 can be used with any other suitable device and any other suitable convolutional neural network architecture.

[0072] During training process 500, server 106 receives multiple initial image frames of the scene, including a short noisy frame 501 and multiple target frames 511-513. The short noisy frame 501 represents... Figure 2input image frames 201 (also short noise frames) that are similar to the input frames. The target frames 511-513 include a crisp-clear frame 511, a normal frame 512, and a long frame 513. The target frames 511-513 correspond to Figure 2 output image frames 211-213 and represent the "targets" for the SIME network 205 when trained using the training process 500.

[0073] In some embodiments, the capture of the initial image frames 501 and 511-513 is performed using a camera on a tripod and triggered with a voice command or other wireless command to avoid touching the camera (and thus moving the camera slightly during capture). This ensures that the image frames 501 and 511-513 are as aligned as possible. In some embodiments, the crisp-clear frame 511 is obtained by darkening the normal frame 512 to the same exposure level as the short noise frame 501.

[0074] To enhance training, the server 106 can perform a data augmentation function 502 on the image frames 501 and 511-513. The data augmentation function 502 increases the size of the set of training data by generating additional images of the scene in addition to the original image frames 501 and 511-513, which can yield better training results overall. For example, the data augmentation function 502 can include transforming the original image frames in different ways, such as flipping the initial image frames 501 and 511-513 along the x and / or y axes, rotating the initial image frames 501 and 511-513 by one or more angles (e.g., 90º, 180º, or 270º), and the like. The transformed images can be used in the network training process 500 along with the initial image frames 501 and 511-513 to create a larger data volume of presentations.

[0075] Once the initial image frames 501 and 511-513 (and any additional image training data obtained through the data augmentation function 502) are obtained, the server 106 trains the SIME network 205 by operating in an iterative manner to generate output images that are similar to the target image frames 511-513. During each iteration in the training process 500, the SIME network 205 generates an image frame 508. Ideally, over multiple iterations in the training process 500, the SIME network 205 improves the generation of image frames 508 overall and progresses toward generating the target image frames 511-513.

[0076] For each iteration in the training process 500, the server 106 may execute a loss calculation function 504, which computes the total loss function of the SIME network 205. This loss helps guide the updating of the weights of the SIME network 205, for example, via gradient descent techniques. For instance, the server 106 may compute the total loss function as a weighted sum of two component loss functions: an L1 loss and a multi-scale structural similarity measure (MSSIM) loss. L1 loss can be used to ensure color accuracy by penalizing color mismatches between target frames 511-513 and image frames 508 generated during each training iteration. Specifically, the L1 loss is the L1 norm (or Manhattan distance) between target frames 511-513 and image frames 508 generated during each training iteration. The MSSIM loss is used to ensure detail preservation in the output by penalizing detail mismatches between target frames 511-513 and image frames 508 generated during each training iteration. The total loss function can be represented as a linear combination of the two component loss functions to ensure that image frame 508 and target frames 511-513 match in both color and detail.

[0077] Based on the calculated loss, electronic device 101 adjusts SIME network 205 by executing weight update function 506 to update the weights used by SIME network 205. For example, server 106 can change the weights used in convolutional layers 311 and 314 or other parameters of SIME network 205. Once the updated weights are determined, server 106 performs another training iteration on SIME network 205. The overall goal of training process 500 is to reduce or minimize the value of the loss function.

[0078] although Figure 5 An example of the process 500 for training the SIME network 205 is shown, but more details can be found elsewhere. Figure 5 Various modifications can be made. For example, the SIME network 205 can be trained in any other suitable manner, which may or may not involve the use of various features such as data augmentation and combined loss functions.

[0079] Figure 6 An example process 600 for generating a ghost-free HDR frame using multiple output image frames generated by a SIME network 205, according to this disclosure, is shown. For ease of illustration, process 600 is described as involving using... Figure 1 Electronic devices 101 and Figure 2 and Figure 3 The SIME network 205 is shown. However, process 600 can be used with any other suitable device and any other suitable convolutional neural network architecture.

[0080] like Figure 6As shown, the electronic device 101 receives a single input image frame 201 and uses the SIME network 205 to generate multiple output image frames 211-213 by using the above-described process. The electronic device 101 inputs the multiple output image frames 211-213 to an HDR blending operation 602, which processes the output image frames 211-213 to generate an HDR image 604. The HDR blending operation 602 applies an HDR blending algorithm to the output image frames 211-213 to generate an HDR image 604 with reduced or minimal artifacts due to the perfect alignment of the output image frames 211-213.

[0081] The HDR blending operation 602 can use any suitable technique to generate the HDR image 604 based on the multiple image frames 211-213, such as by blending the image frames 211-213. For example, one technique for image blending compares and combines individual pixels in the same location of different image frames. Repeating this operation over all pixels of the output image frames 211-213 results in the generation of the HDR image 604. Since the image frames 211-213 are aligned, the HDR blending operation 602 only needs to analyze exposure differences between the output image frames 211-213. This avoids complex (and unreliable) deghosting mechanisms that would be necessary to analyze both exposure and motion differences together (as is typically necessary when image frames are not aligned). Achieving the HDR effect in this way ensures that there is no fundamental conflict between exposure differences and motion differences, providing a reliable and artifact-free HDR result. Note, however, that there are multiple possible techniques for blending image frames, and the HDR blending operation 602 can support any suitable technique or techniques for combining image frames.

[0082] While it can be assumed that a system using a single-input HDR network can achieve similar results as process 600 (combining SIME network 205 and HDR blending operation 602), such a system inherently has some fundamental deficiencies. For example, a single-input HDR network is closed, and thus cannot further improve HDR quality (based on specific image details) by additional captured frames. As a result, the HDR quality achieved by such a network is limited by the quality of its HDR ground truth or target. Obtaining realistic HDR targets for training a single-input HDR network is expensive and does not scale well for productization. For example, the HDR target would need to be obtained by a high dynamic range sensor, which can not be feasible on a handheld device. Also, to obtain the HDR target by a conventional HDR blending algorithm, this becomes a limiting factor. Furthermore, the process of synthesizing low dynamic range (LDR) inputs from HDR images is limited by the synthesis model. In contrast, obtaining realistic high-quality training data for SIME network 205 as discussed above is simple. In addition, the multi-exposure approach of process 600 provides an open system with the potential to reuse the HDR backend (or extend its use) to improve quality.

[0083] While Figure 6 One example of a process 600 for producing artifact-free HDR frames using multiple output image frames generated by SIME network 205 is shown, various changes can be made. Figure 6 For example, SIME network 205 can generate two or more than three image frames.

[0084] As discussed above, processes 200 and 600 use one input image frame 201. However, in some cases, input image frame 201 can not contain much fine detail. To address this possibility, some embodiments of the present disclosure provide a process that can use more than one input image frame, such that the additional image frames can supplement image frame 201. As discussed below, such a process can focus on regions of little or no motion of the short image to minimize the possibility of blur.

[0085] Figures 7A-7D Another example process 700 for producing artifact-free HDR frames using multiple output image frames generated by SIME network 205 in accordance with the present disclosure is shown. Process 700 leverages one or more normal (medium exposure) frames 702 to augment SIME network 205 to improve the quality and detail of the HDR image. Figure 7A An overview of process 700 is shown, while Figures 7B-7D Certain operations of process 700 are shown in more detail. For ease of illustration, process 700 is described as involving the use of Figure 1 electronic device 101 and Figure 2and Figure 3 The process 700 can be used with any other suitable device and any other suitable convolutional neural network architecture.

[0086] As Figure 7A As shown, the electronic device 101 receives a plurality of input image frames 201 and 702 captured with at least one image sensor. The input image frames 201 can be short image frames (denoted with “S” in the figure), and the input image frames 702 can be medium exposure image frames (denoted with “M”) captured with the same sensor as the image frames 201 or a different image sensor. Because the input image frames 702 are brighter than the short frames 201, the input image frames 702 include additional details that can be used to enhance the final HDR frame. Although Figure 7A One medium exposure image frame 702 is depicted, but this is merely an example, and other embodiments using more than one medium exposure image frame are possible and within the scope of the present disclosure.

[0087] The electronic device 101 takes the short input image frames 201 and performs the SIME network 205 to generate a plurality of output image frames 211-213 as described above. The generated output image frames 211-213 include a short clear image frame 211 (denoted with “S”), a normal image frame 212 (possibly exhibiting medium exposure, denoted with “M’”), and a long image frame 213 (denoted with “L”). The electronic device 101 takes the normal image frame 212 from the SIME network 205 and the medium exposure input image frame 702 and performs a motion blur reduction (MBR) deghosting operation 715. The MBR deghosting operation 715 determines and analyzes motion statistics between the normal image frame 212 and the input image frame 702, with the normal image frame 212 being treated as a reference frame. Based on the motion statistics, the MBR deghosting operation 715 generates one or more motion maps 718. The motion maps (also referred to as deghosting maps) generally identify regions in the image frames where motion is occurring and should be removed, thereby identifying desired levels of motion and noise in the image frames. Note that there are a variety of possible techniques for image deghosting, and the MBR deghosting operation 715 can support any one or more suitable techniques for image deghosting. Generally, because the normal image frame 212 and the input image frame 702 can be at the same or close to the same exposure level, the risk of ghosting artifacts is minimal, as deghosting at the same exposure level will be very reliable.

[0088] After the MBR deghosting operation 715, the electronic device 101 performs a plurality of blending operations, including an MBR blending operation 720 and an HDR blending operation 725. Figure 7B Example details of the MBR blending operation 720 are shown, while Figure 7CExample details of the HDR blending operation 725 are shown.

[0089] As Figure 7B shown, the MBR blending operation 720 performs a weighted combination of the normal image frames 212 (reference images) generated by the SIME network 205 and the input image frames 702 (non-reference images) using the motion maps 718 from the MBR deghosting operation 715. The MBR blending operation 720 is performed to introduce details from well-exposed static regions of the captured input image frames 702 into the final HDR image frames. In some embodiments, each motion map 718 can include scalar values, each scalar value identifying a weight to be applied to a respective pixel value in the associated image frame. For example, the pixels in the input image frames 702 can be scaled by the weights of the motion maps 718 using a multiplication function 721. The MBR blending operation 720 can weight and combine the pixels in the same locations of different image frames based on a blending map of the different image frames. Repeating this operation over all pixels of the image frames results in the generation of the medium-exposure MBR blended image 722. The combination of the normal image frames 212 and the input image frames 702 can use any suitable combination technique, e.g., a summation operation 723. Note that there are multiple possible techniques for MBR blended image frames, and the MBR blending operation 720 can support any suitable technique or techniques for MBR image blending. Also, although the MBR blending operation 720 is shown as being performed on the input image frames 702, the MBR blending operation 720 can be performed on any suitable set of image frames, e.g., the normal image frames 212, the input image frames 702, or any other suitable set of image frames. Figure 7B Although N non-reference images are shown, it is understood that N can equal 1, meaning that there is only one non-reference image 702 and one motion map 718.

[0090] As Figure 7C shown, the HDR blending operation 725 performs a pixel-by-pixel exposure analysis of the medium-exposure MBR blended image 722 (output from the MBR blending operation 720), the short clear image frame 211, and the long image frame 213 (output from the SIME network 205) using pixel-by-pixel addition 727. The MBR blended image 722 is used as a reference in the HDR blending operation 725. The HDR blending operation 725 is performed to achieve an HDR effect, e.g., by brightening darker regions and recovering high-light regions. The HDR blending operation 725 results in the generation of the HDR blended image 728.

[0091] In some embodiments, the HDR blending operation 725 performs a weighted combination of the frames 211 and 213 using an overexposed map 731 and an underexposed map 732. The weighting operation can be performed using a pixel-by-pixel multiplication function 726 or any other suitable weighting function. Figure 7D An example process for generating the underexposed map 732 and the overexposed map 731 using a luminance extraction function 736 is shown. As Figure 7DAs shown, the underexposed illustration 732 is generated by thresholding the MBR blended image 722 on sufficiently dark pixels. Here, the luminance extraction function 736 uses a formula 737 that takes into account an underexposure threshold μ U Similarly, the overexposed illustration 731 is generated by thresholding the MBR blended image 722 on sufficiently bright pixels. Here, the luminance extraction function 736 uses a formula 738 that takes into account an overexposure threshold μ O Note that there are multiple possible techniques for HDR blended image frames, and the HDR blending operation 725 can support any one or more suitable techniques for HDR image blending.

[0092] After the HDR blending operation 725, the electronic device 101 can perform a number of post-processing operations using the HDR blended image 728. In this example, the HDR blended image 728 undergoes a tone mapping operation 730 and a sharpening operation 735. The tone mapping operation 730 is generally operable to apply a global tone mapping curve to the HDR blended image 728 in order to brighten dark areas and increase image contrast in the HDR blended image 728. The sharpening operation 735 is generally operable to enhance edges in the HDR blended image 728, such as with a 2D high-pass filter or any other suitable filter. Various techniques for tone mapping and sharpening are known in the art. The output of the process 700 is at least one final image 738. The final image 738 generally represents a blend of the captured image frames 201, 702 after processing. The final image 738 is sent to an image signal processor (ISP) 740 and out of the Bayer domain.

[0093] As noted above, the process 700 combines the SIME network 205 with an HDR pipeline to achieve better quality in both detail and artifacts. Detail in normal image frames 212 from the SIME network 205 is improved by blending with one or more input image frames 702. The blending methods employed in the process 700 include the MBR de-artifacting operation 715 and the MBR blending operation 720. As discussed above, there is minimal risk of artifacting since de-artifacting performs well at the same exposure level. The HDR blending operation 725 is simple here since the SIME network 205 is used, which would involve complex de-artifacting to handle both exposure and motion differences in some conventional systems.

[0094] Although Figures 7A-7D The process 700 shown illustrates one example for producing an artifact-free HDR frame using a number of output image frames generated by the SIME network 205, various changes can be made Figures 7A-7D For example, the SIME network 205 can generate two or more than three image frames.

[0095] Figure 8 Another example process 800 according to this disclosure is shown for generating ghost-free HDR frames using multiple output image frames generated by the SIME network 205. For ease of illustration, process 800 is described as involving using Figure 1 Electronic devices 101 and Figure 2 and Figure 3 The SIME network 205 is shown. However, process 800 can be used with any other suitable device and any other suitable convolutional neural network architecture.

[0096] like Figure 8 As shown, electronic device 101 receives multiple image frames 201 and 702 captured by a camera. Input image frame 201 is a short image frame, and input image frame 702 is a medium-length image frame. Electronic device 101 uses the short input image frame 201 and generates multiple output image frames 211-213 using SIME network 205. In process 800, with Figures 7A-7D The corresponding operation of process 700 is similar; image capture and image processing by SIME network 205 are performed in the Bayer domain. Therefore, output image frames 211-213 and capture frames 201 and 702 are Bayer frames.

[0097] As mentioned above, process 700 is characterized by HDR processing performed in the same Bayer domain. In contrast, Figure 8 The process 800 is characterized by HDR processing performed in the YUV domain. For example... Figure 8 As shown, output image frames 211-213 and capture frames 201, 702 are transmitted to the ISP 808 and converted into image frames in the YUV domain. Then, the electronic device 101 uses the image frames 201, 702, and 211-213 in the YUV domain to perform HDR processing operations. Specifically, the HDR processing includes MBR deghosting operation 815, MBR blending operation 820, HDR blending operation 825, and post-processing operations that may include tone mapping operation 830 and sharpening operation 835. These operations 815-835 are similar to the corresponding operations 715-735 of the above-described process 700, but operations 815-835 are performed in the YUV domain instead of the Bayer domain.

[0098] although Figure 8 An example of a process 800 for generating ghost-free HDR frames using multiple output image frames generated by a SIME network is shown, but further details are possible. Figure 8 Various modifications can be made. For example, the SIME network 205 can generate two or more image frames. Furthermore, the ISP 808 can convert Bayer domain data into other types of image data, such as RGB data.

[0099] Figure 9 An example process 900 for changing an image domain using a SIME network 205 according to this disclosure is shown. For ease of illustration, process 900 is described as involving using Figure 1 Electronic device 101. However, process 900 can be used with any other suitable electronic device and can be used in any suitable system.

[0100] like Figure 9 As shown, electronic device 101 receives multiple image frames 201 and 702. Input image frame 201 is a short image frame, and input image frame 702 includes one or more medium-sized image frames (denoted as M). i (where i is greater than or equal to 1). Frames 201 and 702 are in the raw format. Electronic device 101 takes input image frames 201 and 702 and provides these frames to ISP 904. ISP 904 processes input image frames 201 and 702 and converts these image frames from the raw domain to the corresponding image frames 911-912 in the YUV domain.

[0101] The electronic device 101 also employs a short input image frame 201 and uses a SIME network 205 to generate one or more output image frames in the YUV domain, including a medium-exposure output image frame. In contrast to some embodiments where the SIME network 205 does not alter the domain of the processed image, Figure 9 The SIME network 205 in the image transforms the domain of the short input image frame 201 from the original (Bayer) domain to the YUV domain. Additionally, the SIME network 205 can perform other image processing operations, including demosaicing, noise removal, and color and tone correction. After processing in the SIME network 205, the electronic device 101 can perform one or more post-processing operations 907 on the medium-exposure output image frame, such as dithering, adding synthetic noise, histogram matching, local tone mapping, and other image processing operations. The output of the post-processing operation 907 includes a medium YUV frame 915 that substantially matches the medium input image frame 912 in brightness, color, and hue.

[0102] Then, the electronic device 101 uses a medium YUV frame 915 as a reference frame in multiple mixing operations, including MBR mixing operation 920 and HDR mixing operation 925. Figure 7B Similar to the MBR blending operation 720, the MBR blending operation 920 takes the moving regions from the medium YUV frame 915 and the well-exposed static regions from the medium input image frame 912, and blends these images to introduce details from the static regions. Then, the HDR blending operation 925 recovers the saturated regions from the short input image frame 911 from the ISP 904. The output of process 900 includes at least one final image 930.

[0103] althoughFigure 9 One example of a process 900 for changing an image domain using a SIME network is shown, but various changes can be made Figure 9 For example, the SIME network 205 can generate two or more than three image frames. Also, data in other domains besides the YUV domain can be used, such as when RGB data is used.

[0104] Figure 10 Another example process 1000 for changing an image domain using a SIME network 205 according to the present disclosure is shown. For ease of illustration, the process 1000 is described as involving the use of the electronic device 101. However, the process 1000 can be used with any other suitable electronic device and can be used in any suitable system. Figure 1

[0105] The process 1000 includes many components and operations that are the same or similar to the corresponding components and operations of the process 900 of the electronic device 101. As Figure 9 shown, the electronic device 101 receives a plurality of image frames 201, 702, where the input image frames 201 are short image frames and the input image frames 702 include one or more medium image frames. The frames 201, 702 are in a raw format. Figure 10 The electronic device 101 takes the short input image frames 201 and uses a first SIME network 205a to generate one or more output image frames in the YUV domain that include a medium exposure output image frame. This is similar to the operations performed by the SIME network 205 in

[0106] The electronic device 101 also takes the medium input image frames 702 and performs a second SIME network 205b to generate one or more output frames in the YUV domain that include a medium exposure output image frame. The SIME networks 205a-205b can also perform other image processing operations, including demosaicing, noise removal, and color and tone correction. After processing in the SIME networks 205a-205b, the electronic device 101 performs respective post-processing operations 1007-1008 on the medium exposure output image frames, which can include dithering, adding synthetic noise, histogram matching, local tone mapping, or other image processing operations. The output of the post-processing operation 1008 includes a first medium YUV frame 1010. The output of the post-processing operation 1007 includes a short frame 1011 and a second medium YUV frame 1015 that substantially matches the first medium YUV frame 1010 in brightness, color, and tone. Figure 9

[0107] ​​Then, the electronic device 101 uses a second intermediate YUV frame 1015 as a reference frame in multiple mixing operations, including MBR mixing operation 1020 and HDR mixing operation 1025. These operations are related to... Figure 9 The corresponding operations are the same or similar. The output of process 1000 includes at least one final image 1030.

[0108] although Figure 10 An example of a process 1000 for changing an image domain using a SIME network is shown, but it is possible to modify it further. Figure 10 Various modifications can be made. For example, each SIME network 205a-205b can generate two or more image frames. Furthermore, data from domains other than YUV can be used, such as when using RGB data. Additionally, while process 1000 depicts the use of multiple SIME networks 205a-205b, other embodiments may use a single SIME network that processes different data sequentially (e.g., once for short input image frame 201 and once for medium input image frame 702).

[0109] Figure 11A , Figure 11B , Figure 12A and Figure 12B Examples of benefits that can be achieved using one or more embodiments of this disclosure are shown. Figure 11A and Figure 11B A comparison is depicted between image 1101, captured using a conventional image sensor, and image 1102, captured using one of the disclosed embodiments described above, of the same daytime scene. Figure 11A In the image 1101, the image was captured and processed using traditional HDR operations. (As per...) Figure 11A It is evident that the HDR result for image 1101 is poor, as it includes objects that are too dark. In contrast, Figure 11B Image 1102 was captured and processed using the SIME network described above, along with MBR and HDR blending and deghosting operations. Compared to image 1101, the resulting image 1102 provides superior HDR results and also exhibits good MBR performance.

[0110] Similarly, Figure 12A and Figure 12B A comparison is depicted between image 1201 of a night scene captured using a conventional image sensor and image 1202 of the same night scene captured using one of the embodiments disclosed above. Figure 12A In the image, image 1201 was captured and processed using a long exposure. (As per...) Figure 12A It is clearly visible that the result of image 1201 is poor, as it includes a large amount of noise (as can be clearly seen from the grainy quality of image 1201). In contrast,Figure 12B Image 1202 in the image was captured and processed using a SIME network for processing short exposure frames. (As in...) Figure 12B It is clearly visible that image 1202 is clearer and exhibits very little noise.

[0111] although Figure 11A , Figure 11B , Figure 12A and Figure 12B Examples of the benefits that can be achieved are shown, but various modifications can be made to these figures. For example, images of multiple scenes can be captured under different lighting conditions, and these figures do not limit the scope of this disclosure. These figures are intended only to illustrate examples of the types of benefits that can be obtained using the techniques described above.

[0112] Figure 13 An example method 1300 for generating multi-exposure frames using a SIME network according to this disclosure is shown. For ease of illustration, Figure 13 The method 1300 shown is described as involving the use of Figure 1 Electronic device 101. However, Figure 13 The method 1300 shown can be used with any other suitable electronic device and can be used in any suitable system.

[0113] like Figure 13 As shown, in step 1302, at least one image sensor of the electronic device is used to obtain a first image frame of the scene. This may, for example, include the processor 120 of the electronic device 101 receiving a capture request and having a camera (sensor 180) capture an input image frame 201. In some embodiments, the first image frame is a short-exposure image frame.

[0114] In step 1304, a convolutional neural network is used to generate a plurality of second image frames simulated to have different exposures from the first image frame. This may include, for example, the processor 120 of the electronic device 101 executing the SIME network 205 to generate a plurality of output image frames 211-213. As a result of the processing performed by the SIME network 205, the output image frames 211-213 may be substantially or perfectly aligned with the input image frame 201. Thus, one or more objects in the scene in each second image frame are aligned with one or more corresponding objects in the scene of at least another second image frame and also with one or more corresponding objects in the scene of the first image frame.

[0115] In some embodiments, the second image frames include a first generated image frame with normal exposure, a second generated image frame with a second exposure that is longer than the normal exposure, and a third generated image frame with a third exposure that is shorter than the normal exposure. Also, in some embodiments, the convolutional neural network in the SIME network 205 performs operations in a common encoder path for all of the plurality of second image frames, and performs different operations in separate decoder paths for each of the plurality of second image frames. As described above, each decoder path can include at least one decoder stage that includes a resize or upsample layer, a convolution layer, and a separable convolution layer. The decoder path for the second generated image frame can include more decoder stages than the decoder path for the first generated image frame, and the decoder path for the third generated image frame can include fewer decoder stages than the decoder path for the first generated image frame. Furthermore, in some embodiments, the first image frame is in a raw format, and the convolutional neural network performs operations to convert from raw to YUV, such that the plurality of second image frames are in YUV format.

[0116] At step 1306, a final image of the scene is generated by blending the plurality of second image frames. This can include, for example, the processor 120 of the electronic device 101 performing an HDR blend operation to generate a single frame HDR image. This can also include or alternatively include the processor 120 performing an MBR deghosting operation, an MBR blend operation, and an HDR blend operation to generate a single frame HDR image.

[0117] At step 1308, the final image of the scene is stored, output, or used in some manner. This can include, for example, the processor 120 of the electronic device 101 displaying the final image of the scene on the display 160 of the electronic device 101. This can also include the processor 120 of the electronic device 101 saving the final image of the scene to a camera roll stored in the memory 130 of the electronic device 101. This can also include the processor 120 of the electronic device 101 attaching the final image of the scene to a text message, email, or other communication to be transmitted from the electronic device 101. Of course, the final image of the scene can be used in any other or additional manner.

[0118] Although Figure 13 one example of a method 1300 for generating multiple exposure frames using a SIME network is shown, various changes can be made to Figure 13 the method 1300. For example, although shown as a series of steps, various steps in Figure 13 the method 1300 could overlap, occur in parallel, occur in a different order, or occur any number of times.

[0119] It should be noted that while the various operations are described as being performed by one or more devices, those operations can be implemented in any suitable manner. For example, each of the functions in the electronic device 101 or the server 106 can be implemented or supported using one or more software applications or other software instructions executed by at least one processor 120 of the electronic device 101 or the server 106. In other embodiments, at least some of the functions in the electronic device 101 or the server 106 can be implemented or supported using dedicated hardware components. In general, the operations of each device can be performed using any suitable hardware or any suitable combination of hardware and software / firmware instructions.

[0120] While the present disclosure has been described with reference to various example embodiments, it will be apparent to those of ordinary skill in the art that various changes and modifications can be suggested by the present disclosure. The present disclosure is intended to encompass all such changes and modifications as falling within the scope of the appended claims.

Claims

1. A method for generating images, comprising: obtaining, using at least one image sensor of an electronic device, a first image frame of a scene based on a short exposure time; generating, using a convolutional neural network, a second image frame, a third image frame, and a fourth image frame from the first image frame, wherein the second image frame, the third image frame, and the fourth image frame each have a different exposure level; obtaining, using the at least one image sensor of the electronic device, a fifth image frame of the scene based on a medium exposure time; generating at least one motion map indicating motion between the fifth image frame and the third image frame having a normal exposure level; generating a first blended image by a weighted combination of the fifth image frame and the third image frame based on the at least one motion map; and generating a second blended image based on the first blended image, the second image frame, and the fourth image frame.

2. The method of claim 1, wherein the third image frame is simulated to have a first exposure according to an exposure time, wherein the fourth image frame is simulated to have a second exposure that is longer than the first exposure, wherein the second image frame is simulated to have a third exposure that is shorter than the first exposure, and wherein the first exposure corresponds to an exposure time that is automatically determined resulting in an image frame having least underexposed or / and overexposed areas.

3. The method of claim 1, wherein the second blended image comprises a high dynamic range (HDR) blended image, the HDR image being based on the first blended image as a reference in an HDR blending operation.

4. The method of claim 2, wherein, the convolutional neural network performs operations in a common encoder path for all of the second image frame, the third image frame, and the fourth image frame, wherein the convolutional neural network performs different operations in respective decoder paths for the second image frame, the third image frame, and the fourth image frame, respectively, and wherein each decoder path comprises at least one decoder stage, the at least one decoder stage comprising an up-sampling layer, a convolution layer, and a separable convolution layer.

5. The method of claim 4, wherein the decoder path for the fourth image frame comprises more decoder stages than the decoder path for the third image frame, and wherein the decoder path for the second image frame comprises fewer decoder stages than the decoder path for the third image frame.

6. The method of claim 1, wherein, the first image frame of the scene is in a raw format, wherein the second image frame, the third image frame, and the fourth image frame are in a YUV format, and wherein the convolutional neural network performs a raw-to-YUV conversion.

7. The method of claim 1, wherein, the convolutional neural network is trained by: obtaining a plurality of initial image frames of a scene, the plurality of initial image frames comprising a plurality of target image frames; generating an output image; computing a loss function indicating a difference between the target image frames and the generated output image; updating weights associated with parameters of the convolutional neural network based on the loss function; and repeating the generating operation, the computing operation, and the updating operation one or more times in an iterative manner.

8. An electronic device, comprising: at least one image sensor; and at least one processing device configured to: obtain, using the at least one image sensor, a first image frame of a scene based on a short exposure time; generate, using a convolutional neural network, a second image frame, a third image frame, and a fourth image frame from the first image frame, wherein the second image frame, the third image frame, and the fourth image frame each have a different exposure level; obtain, using the at least one image sensor, a fifth image frame of the scene based on a medium exposure time; and generate at least one motion map indicating motion between the fifth image frame and the third image frame having a normal exposure level; generate a first blended image by weightedly combining the fifth image frame and the third image frame based on the at least one motion map; and generate a second blended image based on the first blended image, the second image frame, and the fourth image frame.

9. The electronic device of claim 8, wherein the third image frame is simulated to have a first exposure according to an exposure time, wherein the fourth image frame is simulated to have a second exposure that is longer than the first exposure, wherein the second image frame is simulated to have a third exposure that is shorter than the first exposure, and wherein the first exposure corresponds to an exposure time that is automatically determined to result in an image frame having least underexposed or / and overexposed areas.

10. The electronic device of claim 8, wherein the second blended image comprises a high dynamic range (HDR) blended image that is based on the first blended image as a reference in an HDR blending operation.

11. The electronic device of claim 9, wherein the convolutional neural network is configured to perform operations in a common encoder path for all of the second image frame, the third image frame, and the fourth image frame, wherein the convolutional neural network is configured to perform different operations in respective decoder paths for the second image frame, the third image frame, and the fourth image frame, respectively, and wherein each decoder path comprises at least one decoder stage comprising an up-sampling layer, a convolution layer, and a separable convolution layer.

12. The electronic device of claim 11, wherein, the decoder path for the fourth image frame comprises more decoder stages than the decoder path for the third image frame, and wherein the decoder path for the second image frame comprises fewer decoder stages than the decoder path for the third image frame.

13. The electronic device of claim 8, wherein the first image frame of the scene is in a raw format, wherein the second image frame, the third image frame, and the fourth image frame are in a YUV format, and wherein the convolutional neural network is configured to perform a raw-to-YUV conversion.

14. The electronic device of claim 8, wherein, the convolutional neural network is trained by: obtaining a plurality of initial image frames of a scene, the plurality of initial image frames including a plurality of target image frames; generating an output image similar to the target image frames; computing a loss function indicative of a difference between the target image frames and the generated output image; updating weights associated with parameters of the convolutional neural network based on the loss function; and repeating the generating operation, the computing operation, and the updating operation one or more times in an iterative manner.

15. A non-transitory machine-readable medium containing instructions that, when executed, cause at least one processor of an electronic device to: obtain, using at least one image sensor of the electronic device, a first image frame of a scene based on a short exposure time; generate, using a convolutional neural network, a second image frame, a third image frame, and a fourth image frame from the first image frame, wherein the second image frame, the third image frame, and the fourth image frame each have a different exposure level; obtain, using the at least one image sensor of the electronic device, a fifth image frame of the scene based on a medium exposure time; generate at least one motion map indicative of motion between the fifth image frame and the third image frame having a normal exposure level; generate a first blended image by weightedly combining the fifth image frame and the third image frame based on the at least one motion map; and generate a second blended image based on the first blended image, the second image frame, and the fourth image frame.

16. The non-transitory machine-readable medium of claim 15, the third image frame is simulated to have a first exposure according to an exposure time, wherein, wherein the fourth image frame is simulated to have a second exposure longer than the first exposure, wherein the second image frame is simulated to have a third exposure shorter than the first exposure, and wherein the first exposure corresponds to an exposure time determined automatically resulting in the image frame having least underexposed or / and overexposed areas. the convolutional neural network is configured to perform operations in a common encoder path for all of the second image frame, the third image frame, and the fourth image frame, 17. The non-transitory machine-readable medium of claim 16, wherein, wherein the convolutional neural network is configured to perform different operations in respective decoder paths for the second image frame, the third image frame, and the fourth image frame, respectively, and wherein each decoder path includes at least one decoder stage including an up-sampling layer, a convolution layer, and a separable convolution layer.

18. The non-transitory machine-readable medium of claim 17, the decoder path for the fourth image frame includes more decoder stages than the decoder path for the third image frame, and wherein wherein the decoder path for the second image frame includes fewer decoder stages than the decoder path for the third image frame. ​

Images