Increasing ai robustness using GAN selector

Abstract

Deep neural networks (DNNs) can denoise corrupted inputs that result from natural corruptions, so as improve the performance and trustworthiness of the DNNs. In particular, for example, based on the noise type in a given image, a generative adversarial network (GAN) selector can be trained to select one or more GANs from a plurality of GANs. The selected GANs can be selected because they are the best GANs to denoise the particular image. The selected GANs can denoise the image, so that objects in the corrupted input can be detected and identified.

Description

Docket No. 202306035 INCREASING AI ROBUSTNESS USING GAN SELECTOR BACKGROUND

[0001] Object detection and classification is an increasingly important task in ensuring the safety of autonomous machines, such as the autonomous vehicles, trains, robots, or the like. Artificial intelligence (AI) models can be deployed in the field for object detection and classification, but such models are generally not functioning in their ideal trained environment. For example, AI models or deep neural networks (DDNs) can encounter various natural corruptions (e.g., fog, snow, etc.) in the physical environment that are underrepresented in their training environment. Such corruptions can add noise to images that AI models process to detect objects. Noise can cause DNNs to misclassify objects or otherwise lead to hazardous situations. BRIEF SUMMARY

[0002] Embodiments of the invention address and overcome one or more of the described- herein shortcomings or technical problems by providing methods, systems, and apparatuses for increasing the robustness of artificial intelligence (AI) systems, so as to better classify objects with natural corruptions. In particular, for example, deep neural networks (DNNs) can denoise corrupted input that results from natural corruptions, so as improve the performance and trustworthiness of the DNNs.

[0003] In an example aspect, an object classification computer system includes a camera configured to capture an image of a physical environment. The image can include an object and noise, so as to define a naturally corrupted image. The system can further include a noise classifier configured to determine one or more noise types of the noise in the image. The system can further include a plurality of generative adversarial networks (GANs) configured to denoise corrupted images. The system can further include a generative adversarial network (GAN) selector configured to, based on the one or more noise types, select one or more GANs from the plurality of GANs, so as to define one or more selected GANs. The one or more selected GANs can be configured to denoise the image, so as to generate respective denoised inputs. Furthermore, a deep neural network (DNN) can be configured to generate an uncertainty for each of the respective denoised inputs. Based on the denoised input associatedDocket No. 202306035 with the uncertainty that is lower than the uncertainty of the other denoised inputs, the system can detect the object in the naturally corrupted image. The noise can be defined by snow, fog, rain, glare, or the like.

[0004] In another example aspect, the system can train the GAN selector with first images that the plurality of GANs misclassified, so as to define misclassified data. The GAN selector can be further trained with second images that at least one of the plurality of GANs correctly classified, so as to define correctly classified data. The first images can define a specific object and a specific noise type, and the second images can also define the specific object and the specific noise type. In various examples, the first images can define a first noise intensity that is less than a second noise intensity defined by the second images, such that each of the first images are equivalent to respective second images except for the respective noise intensities. BRIEF DESCRIPTION OF THE SEVERAL VIEWS OF THE DRAWINGS

[0005] The foregoing and other aspects of the present invention are best understood from the following detailed description when read in connection with the accompanying drawings. For the purpose of illustrating the invention, there is shown in the drawings embodiments that are presently preferred, it being understood, however, that the invention is not limited to the specific instrumentalities disclosed. Included in the drawings are the following Figures:

[0006] FIG. 1 shows an example system that includes a train and at least one camera in communication with a central computer system, in accordance with an example embodiment.

[0007] FIG. 2 illustrates an example object detection computer system that can be defined by the central computer system, in accordance with an example embodiment.

[0008] FIG. 3 shows example neural networks configured to train a generative adversarial network (GAN) selector neural network, wherein the neural networks including a GAN selector can be defined by the object detection computer system.

[0009] FIG. 4 illustrates the objection classification computer system including various neural networks configured to detect and identify objects in naturally corrupted images.

[0010] FIG. 5 illustrates a computing environment within which embodiments of the disclosure may be implemented.Docket No. 202306035 DETAILED DESCRIPTION

[0011] Referring now to FIG. 1, an example autonomous system 100 can identify objects related to operations of an autonomous machine such as autonomous vehicle or train 102. For purposes of example, the autonomous machine is illustrated as an autonomous train, though it will be understood that the autonomous machine can define various other machines that perform object detection or classification, such as autonomous cars, automated guided vehicles (AGVs), automated mobile robots, or the like, and all such machines are contemplated as being within the scope of this disclosure. By way of further example, embodiments described herein can be implemented in various manufacturing stations that use a camera, for instance stations that use a camera to control quality of a product (e.g., detecting dents in metal, damage to packaging, etc.). The system 100 can include the train 102 that includes one or more train cars 105, for instance a first train car 105a, a second train car 105b, and a third train car 105c, configured to run on a track 115. It will be understood that the illustrated system 100 including the train 102 are simplified for purposes of example.

[0012] Still referring to FIG. 1, the system 100, for instance the train 102, can include one or more cameras or sensors, for instance RGBD cameras 118, configured to detect or record objects within the physical environment, for instance on, or adjacent to, the track 115. In some cases, the one or more cameras of the system 100, can include one or more standard two- dimensional (2D) cameras that can record or capture images (e.g., RGB images or depth images) from different viewpoints. Those images can be used to construct 3D images. For example, and without limitation, 2D cameras can be mounted to the front of the train 105 or overhead the track 115, so as to capture images from perspectives along the path of the train 102. It will be understood that that the number of sensors in a given system 100 may vary as desired.

[0013] One or more of the train cars 105, for instance each train car, can define a computer system 108 configured to collect data from the cameras 118. In some cases, each computer system 108 can be communicatively coupled to a central computing system 110 that be located on the train 102 or separate from the train 102, such as within a local area network (LAN) that is communicatively coupled to the train 102. A train control system 112 can also be located on the train 102, for instance within the first train car. In some examples, the central computing system 110 defines the train control system 112. Alternatively, the central computing system 110 can be separately located from, and communicatively computed to, the train control systemDocket No. 202306035 (TCS) 112. The train control system 112 can perform various functions related to controlling the train operation. With respect to object detection, the train control system 112 can receive detection messages from the computer systems 108 and / or the central computing system 110.

[0014] With continuing reference to FIG. 1, the computer systems 108 and the train control system 112 can be connected via communications network 114. The communication network 114 may utilize conventional transmission technologies including, for example, Ethernet and Wi-Fi to facilitate communications between the train cars 105. Each bogie computer system 108 may implement one or more transport layer protocols such as TCP and / or UDP. In some embodiments, the computer system 108 includes functionality that allows the transport protocol to be selected based on real-time requirements or a guaranteed quality of service. For example, for near-real time communications UDP may be used by default, while TCP might be used for communications that have more lax timing requirements but require additional reliability.

[0015] Referring now to FIG. 2, the central computing system 110, the bogie computer system 108, or any combination thereof can define an example object classification computer system 200. In some cases, the object classification computer system 200 can include one or more computing processors configured to process information and control operations of the system 100, in particular the train 102. An autonomous system for operating an autonomous machine within a physical environment can further include a memory for storing modules. The processors can further be configured to execute the modules so as to process information and detect objects based on the information. The object classification system can define one or more neural networks, for instance a generative adversarial network (GAN) 201.

[0016] Still referring to FIG. 2, the GAN 201 can define a first DNN or generator 202 and a second DNN or classifier 204 configured for object recognition. For example, given an input (^), for instance a noisy or corrupted input 206, the generator 202 (^), can generate an output 208 or ^(^), in particular a denoised image that can be input into the classifier 204. The noisy input 206 can represent an original image 210 or (^^^^^) that is naturally corrupted so as to define the corrupted image 206. For instance,or corrupted image 206 can define an image that is blurry (e.g., includes fog, dust, or is otherwise unclear), or an image that was underexposed (e.g., at night or the) or overexposed (e.g., includes sunlight or reflections or is otherwise unusually bright). In various examples, the output 208 of the generator 202 also defines a perturbation 208. An objective of the GAN 201 is to generate a denoised input, ^ + ^(^), after applying the generated perturbation 208 to the noisy input 206. To do so, a sampleDocket No. 202306035 can be created that is similar and belongs to the same data distribution as the original image 210 (^^^^^). Furthermore, the generated denoised output 208 should yield similar predictions as the input 219 (^^^^^), from the classifier 204. Further still, it is recognized herein that the between the original image 210 (^^^^^) and the generated image^ + ^(^) might not to yield similar predictions from the classifier 204.

[0017] For example, when a sample ^ + ^(^) is generated, the Euclidean pixel difference between ^ + ^(^) and the original image 210 (^^^^^) can be represented by δ, where δ → 0. It is recognized herein that in some cases, the value of δ will be greater than 0 such that a pixel difference can exist between the two images. It is further recognized herein that pixel differences, for instance small pixel differences, can cause incorrect predictions or classifications by the classifier 204 To address incorrect predictions, in accordance with various examples, hidden states of the DNN 204 can be analyzed to determine a certainty of the model, and the certainty of the DNN 204 can be included into its loss function. In particular, for example, the difference between hidden layer outputs 212a and 212b of a given original image and a given generated image, respectively can be included in the loss function. Furthermore, to denoise a corrupted image (^), the GAN 201 can generate the perturbation 208 (^(^)) instead of generating a benign input directly. In an example, the benign input can be generated by adding the perturbation 208 with the corrupted input (^). Thus, the total loss function (^^^^^^) can be represented as ^^^^^^= ^^^^+ ||^^ + ^(^)^ − ^^^^^|| + ||^^^(^)+ ^(^) − ^^^(^^^^^)||, where ^^^ represents hidden layer output values 212 of the DNN 204.

[0018] The generator 202 can be trained such that the generated denoised image 208 is nearest to the original image 210 in terms of pixel distance, and the Euclidean difference between hidden state values 212a and 212b of the DNN 204 corresponding to a given original image and denoised image, respectively, is minimum. An objective of the training is to generate the denoised input ^ + ^(^) after applying the generated perturbation 208 to the noisy input 206. The generator 202 can be trained with assistance from the discriminator (DNN) or classifier 204, wherein the classifier 204 can be configured to distinguish between the generated samples and the in-distribution samples. For example, as training continues, the generator 202 can generate increasingly unnoticeable perturbation 208, and the classifier or DNN 204 can be increasingly more accurate in distinguishing in-distribution samples and generated samples. In various examples, after being well trained, the classifier 204 and the generator 202 can reach a Nash Equilibrium, wherein the generated test samples are difficult to distinguish from the in-Docket No. 202306035 distribution samples. Furthermore, with respect to denoising, the certainty of the DNN 204 can be included into the loss function. It is recognized herein that the hidden states of the DNN 204 can be analyzed to determine model certainty. Therefore, in various examples, the differences between hidden layer outputs 212a and 212b of original images and corresponding generated images, respectively, are included in the loss function.

[0019] Referring also to FIG. 3, it is recognized herein that, in some cases, training a single GAN 201 is not sufficient to denoise various, for instance all, types of noises. For example, different noises on benign inputs can result in different effects. By way of further example, the perturbation needed to denoise low contrast images might be similar (but not the same) to the perturbation needed to denoise images affected with glass blur noise. Thus, in accordance with various examples, a plurality of GANs 201 are trained for different noises so that multiple GANs having different architecture and parameters can differently de-noise.

[0020] Referring in particular to FIGs. 3 and 4, the object classification computer system 200 can further include a plurality of GANs 201, for instance a first GAN 201a, a second GAN 201b, and a third GAN 201c, coupled to DNN 302 that is coupled to a GAN selector 304. The object classification system 200 can further include a noise classifier 402 configured to classify noises from input 401. It will be understood that three GANs are illustrated for simplicity, though the system 200 can include an alternative number of GANs, and all such numbers of GANs are contemplated as being within the scope of this disclosure. The GANs 201 can be trained to denoise corrupted inputs, and the GAN selector 304 can be trained and configured to determine how the GANs 201 perform to denoise a given noisy input, so as to select respective GANs 201 based on the input.

[0021] Referring in particular to FIG. 3, the GANs 201 can be trained to denoise corrupted inputs to generate benign inputs from the corrupted inputs. The GANs can be trained with various noisy or corrupted images that define images that are blurry (e.g., include fog, dust, or are otherwise unclear), or defines images that are overexposed (e.g., includes sunlight or reflections or is otherwise unusually bright). In particular, for example, the GANs 201 can be trained on the corrupted images 206, which can define training data that represents various types of corruptions. Furthermore, the loss function can be used to train the GANs 201 as described above. In example, to determine how good each GAN 201 is per noise type, the DNN 302 evaluates the respective closest de-noised image (e.g., an image in which the GAN did not differentiate from the original) from each GAN 201 to determine whether the DNN 302 canDocket No. 202306035 predict the correct classes. By way of example, the system can determine whether the DNN 302 correctly classifies a blurry human in an image as a human. The GAN selector 304 can be trained with the results of the DNN 302 and the corresponding original images modified by different levels of noise, and misclassified images. Thus, each result of the GAN can be captured for each individual level of noise. Consequently, in various examples, multiple GANs can recreate the original image well and differently, which can result in successful classification by the DNN 302.

[0022] Referring in particular to FIG. 4, the noise classifier 402 can be configured to determine which types of noise are in a given image. Based on the type of noise, because the system knows how the GANs 201 perform with respect to noise type, the most appropriate GAN can be selected. When the GAN selector 304 receives the type of noise from the noise classifier 402, the GAN selector can determine and select the “best” one or more (e.g., two) GANs 201 for the noise types(s). Both results of the classifier DNN 302 can be compared by their uncertainties 403 when both have the correct classification predicted.

[0023] In some examples, after the GANs 201 and the noise classifier 402 are trained, the GAN selector 304 is trained. In various examples, the GAN selector 304 defines a convolutional neural network classifier configured to select a given GAN from the plurality of GANs 201 for denoising a specific noise. To train the GAN selector 304, in various example embodiments, certain validation data is used. For example, validation data in addition to, or instead of, denoised inputs is used. It is recognized herein that the DNN of any GAN might correctly classify most, for instance over 99%, of denoised training inputs. Thus, it is further recognized herein that it can be difficult to the train the GAN selector 304 with denoised training inputs.

[0024] In an example, misclassified samples from a specific or first noise type are obtained from the training of the GANs 201, for training the GAN selector 304. The misclassified samples or data can be divided into a first part and a second part. The first and second parts can be equal or otherwise divided into nonequal portions. The first part of the misclassified samples can define a portion of training data or inputs 301 for the GAN selector 304. Images that are classified correctly that correspond to images in the first part of the misclassified data, but represent different noise levels, can also define a portion of the training data 301 for the GAN selector 304. Thus, the training data 301 for the GAN selector 304 can consist of correctly classified images and misclassified images of the same objects or physicalDocket No. 202306035 environment, at different amounts of the same noise type. The second part of the misclassified data can be used for testing the GAN selector 304.

[0025] In an example, the training data can define standard data (e.g., MNIST) that includes various types of noise (e.g., blur, snow, glare, etc.) at various intensities (e.g., light snow, snow rain, heavy snow, etc.). In some cases, the intensities of each noise are grouped into different level, for instance four (4) levels that range from low to high intensity, though it will be understood that additional or alternative noise intensity levels can be used, and all such noise levels are contemplated as being within the scope of this disclosure. In some cases, there are a limited number of images in the dataset that meet the training requirements (e.g., noise types, levels, classes you want to predict, etc.). Thus, the dataset can be split into a larger set (e.g., 70 to 80%) that is used to train the network, and a smaller set (20 to 30%) that can be used for validation of a reasonably good success rate.

[0026] Based on the training data 301, the GANs 201 can generate respective denoised images that define denoised input 303 for the DNN 302. The denoised input 302 can include image labels that indicate the ground truth of what classes are seen in the respective image. Based on the determinations, in some cases, the DNN 302 can generate a 2D array of Boolean labels that indicate whether a given denoised input 302 differs from a respective original input. The inputs 303 that are not correctly denoised by any of the GANs 201 are removed from training data 301, such that the rest of the training data 301 can be input into the multilabel classifier, in particular to the GAN selector 304, so as to train the GAN selector 304.

[0027] Referring in particular to FIG. 4, during an example inference after the GAN selector 304 is trained, a corrupted input 401, for instance a noisy image captured by the camera 118, can be fed to the noise classifier 402. The noise classifier 402 can determine one or more types of noise that are present in the corrupted input. In an example, a multi-label DNN can predict multiple classes (labels). Based on the classifications or determinations that the noise classifier 402 makes, the GAN selector 304 can select one or more of the GANs 201 for denoising the input 401. In various examples, the GANs 201 can include three or more GANs for eight or more different noise types. Thus, the GAN selector 304 might select a single GAN to process the input 401, based on the noise type present in the input 401. Alternatively, the GAN selector 304 might select more than one GAN 201, for instance two GANs, that are the best GANs to process the input 401, based on the one or more noise types present in the input 401. The input 401 is sent to the GANs 201 that are selected by the GAN selector 304. For example, the inputDocket No. 202306035 401 might be sent to the two of the GANs 201 selected by the GAN selector 304, thereby defining selected GANs. The input 401 can be sent to the selected GANs so that the selected GANs can denoise the input 401, so as to define respective denoised inputs for the DNN 302. The respective denoised inputs can be sent to the DNN 302. Based on the denoised inputs, the DNN 302 can generate respective uncertainty scores, or confidence levels, associated with their predictions (classes). The system 200 can compare the uncertainty scores, and select the denoised input that corresponds to the lowest uncertainty score in the DNN 302.

[0028] Thus, the system can determine a classification (class) of an object and, based on the class, control decisions can be made. Referring again to FIG. 1, by way of example, if a given image classification is a human on the track, the system might automatically apply brakes to stop the train before colliding with the human. Alternative, if the given classification is another animal, the system might continue driving the train. Without being bound by theory, this example can illustrate the importance of correct classification. By way of further example, without the object classification system described herein, a human might misclassify the object, for instance due to heavy snow or fog, which can result in fatal injury or other damage. This example also illustrates a need for the correct denoising of the image, so that the most appropriate GAN is selected to make such control decisions with extremely high reliability. Brake applications, among others, illustrate that even an increase in reliability of 1% in object classification can result in hugely beneficial consequences, for instance the saving of lives or money.

[0029] FIG. 5 illustrates an example of a computing environment within which embodiments of the present disclosure may be implemented. A computing environment 500 includes a computer system 510 that may include a communication mechanism such as a system bus 521 or other communication mechanism for communicating information within the computer system 510. The computer system 510 further includes one or more processors 520 coupled with the system bus 521 for processing the information. The central computer system 110 may include, or be coupled to, the one or more processors 520.

[0030] The processors 520 may include one or more central processing units (CPUs), graphical processing units (GPUs), or any other processor known in the art. More generally, a processor as described herein is a device for executing machine-readable instructions stored on a computer readable medium, for performing tasks and may comprise any one or combination of, hardware and firmware. A processor may also comprise memory storing machine-readableDocket No. 202306035 instructions executable for performing tasks. A processor acts upon information by manipulating, analyzing, modifying, converting or transmitting information for use by an executable procedure or an information device, and / or by routing the information to an output device. A processor may use or comprise the capabilities of a computer, controller or microprocessor, for example, and be conditioned using executable instructions to perform special purpose functions not performed by a general purpose computer. A processor may include any type of suitable processing unit including, but not limited to, a central processing unit, a microprocessor, a Reduced Instruction Set Computer (RISC) microprocessor, a Complex Instruction Set Computer (CISC) microprocessor, a microcontroller, an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), a System-on-a-Chip (SoC), a digital signal processor (DSP), and so forth. Further, the processor(s) 520 may have any suitable microarchitecture design that includes any number of constituent components such as, for example, registers, multiplexers, arithmetic logic units, cache controllers for controlling read / write operations to cache memory, branch predictors, or the like. The microarchitecture design of the processor may be capable of supporting any of a variety of instruction sets. A processor may be coupled (electrically and / or as comprising executable components) with any other processor enabling interaction and / or communication there-between. A user interface processor or generator is a known element comprising electronic circuitry or software or a combination of both for generating display images or portions thereof. A user interface comprises one or more display images enabling user interaction with a processor or other device.

[0031] The system bus 521 may include at least one of a system bus, a memory bus, an address bus, or a message bus, and may permit exchange of information (e.g., data (including computer-executable code), signaling, etc.) between various components of the computer system 510. The system bus 521 may include, without limitation, a memory bus or a memory controller, a peripheral bus, an accelerated graphics port, and so forth. The system bus 821 may be associated with any suitable bus architecture including, without limitation, an Industry Standard Architecture (ISA), a Micro Channel Architecture (MCA), an Enhanced ISA (EISA), a Video Electronics Standards Association (VESA) architecture, an Accelerated Graphics Port (AGP) architecture, a Peripheral Component Interconnects (PCI) architecture, a PCI-Express architecture, a Personal Computer Memory Card International Association (PCMCIA) architecture, a Universal Serial Bus (USB) architecture, and so forth.Docket No. 202306035

[0032] Continuing with reference to FIG. 5, the computer system 510 may also include a system memory 530 coupled to the system bus 521 for storing information and instructions to be executed by processors 520. The system memory 530 may include computer readable storage media in the form of volatile and / or nonvolatile memory, such as read only memory (ROM) 531 and / or random access memory (RAM) 532. The RAM 532 may include other dynamic storage device(s) (e.g., dynamic RAM, static RAM, and synchronous DRAM). The ROM 531 may include other static storage device(s) (e.g., programmable ROM, erasable PROM, and electrically erasable PROM). In addition, the system memory 530 may be used for storing temporary variables or other intermediate information during the execution of instructions by the processors 520. A basic input / output system 533 (BIOS) containing the basic routines that help to transfer information between elements within computer system 510, such as during start-up, may be stored in the ROM 531. RAM 532 may contain data and / or program modules that are immediately accessible to and / or presently being operated on by the processors 520. System memory 530 may additionally include, for example, operating system 534, application programs 535, and other program modules 536. Application programs 535 may also include a user portal for development of the application program, allowing input parameters to be entered and modified as necessary.

[0033] The operating system 534 may be loaded into the memory 530 and may provide an interface between other application software executing on the computer system 510 and hardware resources of the computer system 510. More specifically, the operating system 534 may include a set of computer-executable instructions for managing hardware resources of the computer system 510 and for providing common services to other application programs (e.g., managing memory allocation among various application programs). In certain example embodiments, the operating system 534 may control execution of one or more of the program modules depicted as being stored in the data storage 540. The operating system 534 may include any operating system now known or which may be developed in the future including, but not limited to, any server operating system, any mainframe operating system, or any other proprietary or non-proprietary operating system.

[0034] The computer system 510 may also include a disk / media controller 543 coupled to the system bus 521 to control one or more storage devices for storing information and instructions, such as a magnetic hard disk 541 and / or a removable media drive 542 (e.g., floppy disk drive, compact disc drive, tape drive, flash drive, and / or solid state drive). Storage devices 540 mayDocket No. 202306035 be added to the computer system 510 using an appropriate device interface (e.g., a small computer system interface (SCSI), integrated device electronics (IDE), Universal Serial Bus (USB), or FireWire). Storage devices 541, 542 may be external to the computer system 510.

[0035] The computer system 510 may also include a field device interface 565 coupled to the system bus 521 to control a field device 566, such as a device used in a production line. The computer system 510 may include a user input interface or GUI 561, which may comprise one or more input devices, such as a keyboard, touchscreen, tablet and / or a pointing device, for interacting with a computer user and providing information to the processors 520.

[0036] The computer system 510 may perform a portion or all of the processing steps of embodiments of the invention in response to the processors 520 executing one or more sequences of one or more instructions contained in a memory, such as the system memory 530. Such instructions may be read into the system memory 530 from another computer readable medium of storage 540, such as the magnetic hard disk 541 or the removable media drive 542. The magnetic hard disk 541 (or solid state drive) and / or removable media drive 542 may contain one or more data stores and data files used by embodiments of the present disclosure. The data store 540 may include, but are not limited to, databases (e.g., relational, object-oriented, etc.), file systems, flat files, distributed data stores in which data is stored on more than one node of a computer network, peer-to-peer network data stores, or the like. The data stores may store various types of data such as, for example, skill data, sensor data, or any other data generated in accordance with the embodiments of the disclosure. Data store contents and data files may be encrypted to improve security. The processors 520 may also be employed in a multi-processing arrangement to execute the one or more sequences of instructions contained in system memory 530. In alternative embodiments, hard-wired circuitry may be used in place of or in combination with software instructions. Thus, embodiments are not limited to any specific combination of hardware circuitry and software.

[0037] As stated above, the computer system 510 may include at least one computer readable medium or memory for holding instructions programmed according to embodiments of the invention and for containing data structures, tables, records, or other data described herein. The term “computer readable medium” as used herein refers to any medium that participates in providing instructions to the processors 520 for execution. A computer readable medium may take many forms including, but not limited to, non-transitory, non-volatile media, volatile media, and transmission media. Non-limiting examples of non-volatile media include opticalDocket No. 202306035 disks, solid state drives, magnetic disks, and magneto-optical disks, such as magnetic hard disk 541 or removable media drive 542. Non-limiting examples of volatile media include dynamic memory, such as system memory 530. Non-limiting examples of transmission media include coaxial cables, copper wire, and fiber optics, including the wires that make up the system bus 521. Transmission media may also take the form of acoustic or light waves, such as those generated during radio wave and infrared data communications.

[0038] Computer readable medium instructions for carrying out operations of the present disclosure may be assembler instructions, instruction-set-architecture (ISA) instructions, machine instructions, machine dependent instructions, microcode, firmware instructions, state- setting data, or either source code or object code written in any combination of one or more programming languages, including an object oriented programming language such as Smalltalk, C++ or the like, and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The computer readable program instructions may execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server. In the latter scenario, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection may be made to an external computer (for example, through the Internet using an Internet Service Provider). In some embodiments, electronic circuitry including, for example, programmable logic circuitry, field-programmable gate arrays (FPGA), or programmable logic arrays (PLA) may execute the computer readable program instructions by utilizing state information of the computer readable program instructions to personalize the electronic circuitry, in order to perform aspects of the present disclosure.

[0039] Aspects of the present disclosure are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the disclosure. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, may be implemented by computer readable medium instructions.

[0040] The computing environment 500 may further include the computer system 510 operating in a networked environment using logical connections to one or more remoteDocket No. 202306035 computers, such as remote computing device 580. The network interface 570 may enable communication, for example, with other remote devices 580 or systems and / or the storage devices 541, 542 via the network 571. Remote computing device 580 may be a personal computer (laptop or desktop), a mobile device, a server, a router, a network PC, a peer device or other common network node, and typically includes many or all of the elements described above relative to computer system 510. When used in a networking environment, computer system 510 may include modem 572 for establishing communications over a network 571, such as the Internet. Modem 572 may be connected to system bus 521 via user network interface 570, or via another appropriate mechanism.

[0041] Network 571 may be any network or system generally known in the art, including the Internet, an intranet, a local area network (LAN), a wide area network (WAN), a metropolitan area network (MAN), a direct connection or series of connections, a cellular telephone network, or any other network or medium capable of facilitating communication between computer system 510 and other computers (e.g., remote computing device 580). The network 571 may be wired, wireless or a combination thereof. Wired connections may be implemented using Ethernet, Universal Serial Bus (USB), RJ-6, or any other wired connection generally known in the art. Wireless connections may be implemented using Wi-Fi, WiMAX, and Bluetooth, infrared, cellular networks, satellite or any other wireless connection methodology generally known in the art. Additionally, several networks may work alone or in communication with each other to facilitate communication in the network 571.

[0042] It should be appreciated that the program modules, applications, computer-executable instructions, code, or the like depicted in FIG. 5 as being stored in the system memory 530 are merely illustrative and not exhaustive and that processing described as being supported by any particular module may alternatively be distributed across multiple modules or performed by a different module. In addition, various program module(s), script(s), plug-in(s), Application Programming Interface(s) (API(s)), or any other suitable computer-executable code hosted locally on the computer system 510, the remote device 580, and / or hosted on other computing device(s) accessible via one or more of the network(s) 571, may be provided to support functionality provided by the program modules, applications, or computer-executable code depicted in FIG. 5 and / or additional or alternate functionality. Further, functionality may be modularized differently such that processing described as being supported collectively by the collection of program modules depicted in FIG. 5 may be performed by a fewer or greaterDocket No. 202306035 number of modules, or functionality described as being supported by any particular module may be supported, at least in part, by another module. In addition, program modules that support the functionality described herein may form part of one or more applications executable across any number of systems or devices in accordance with any suitable computing model such as, for example, a client-server model, a peer-to-peer model, and so forth. In addition, any of the functionality described as being supported by any of the program modules depicted in FIG. 5 may be implemented, at least partially, in hardware and / or firmware across any number of devices.

[0043] It should further be appreciated that the computer system 510 may include alternate and / or additional hardware, software, or firmware components beyond those described or depicted without departing from the scope of the disclosure. More particularly, it should be appreciated that software, firmware, or hardware components depicted as forming part of the computer system 510 are merely illustrative and that some components may not be present or additional components may be provided in various embodiments. While various illustrative program modules have been depicted and described as software modules stored in system memory 530, it should be appreciated that functionality described as being supported by the program modules may be enabled by any combination of hardware, software, and / or firmware. It should further be appreciated that each of the above-mentioned modules may, in various embodiments, represent a logical partitioning of supported functionality. This logical partitioning is depicted for ease of explanation of the functionality and may not be representative of the structure of software, hardware, and / or firmware for implementing the functionality. Accordingly, it should be appreciated that functionality described as being provided by a particular module may, in various embodiments, be provided at least in part by one or more other modules. Further, one or more depicted modules may not be present in certain embodiments, while in other embodiments, additional modules not depicted may be present and may support at least a portion of the described functionality and / or additional functionality. Moreover, while certain modules may be depicted and described as sub-modules of another module, in certain embodiments, such modules may be provided as independent modules or as sub-modules of other modules.

[0044] Although specific embodiments of the disclosure have been described, one of ordinary skill in the art will recognize that numerous other modifications and alternative embodiments are within the scope of the disclosure. For example, any of the functionalityDocket No. 202306035 and / or processing capabilities described with respect to a particular device or component may be performed by any other device or component. Further, while various illustrative implementations and architectures have been described in accordance with embodiments of the disclosure, one of ordinary skill in the art will appreciate that numerous other modifications to the illustrative implementations and architectures described herein are also within the scope of this disclosure. In addition, it should be appreciated that any operation, element, component, data, or the like described herein as being based on another operation, element, component, data, or the like can be additionally based on one or more other operations, elements, components, data, or the like. Accordingly, the phrase “based on,” or variants thereof, should be interpreted as “based at least in part on.”

[0045] Although embodiments have been described in language specific to structural features and / or methodological acts, it is to be understood that the disclosure is not necessarily limited to the specific features or acts described. Rather, the specific features and acts are disclosed as illustrative forms of implementing the embodiments. Conditional language, such as, among others, “can,” “could,” “might,” or “may,” unless specifically stated otherwise, or otherwise understood within the context as used, is generally intended to convey that certain embodiments could include, while other embodiments do not include, certain features, elements, and / or steps. Thus, such conditional language is not generally intended to imply that features, elements, and / or steps are in any way required for one or more embodiments or that one or more embodiments necessarily include logic for deciding, with or without user input or prompting, whether these features, elements, and / or steps are included or are to be performed in any particular embodiment.

[0046] The flowchart and block diagrams in the Figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present disclosure. In this regard, each block in the flowchart or block diagrams may represent a module, segment, or portion of instructions, which comprises one or more executable instructions for implementing the specified logical function(s). In some alternative implementations, the functions noted in the block may occur out of the order noted in the Figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that each block of the block diagrams and / or flowchart illustration, and combinations of blocksDocket No. 202306035 in the block diagrams and / or flowchart illustration, can be implemented by special purpose hardware-based systems that perform the specified functions or acts or carry out combinations of special purpose hardware and computer instructions.

Claims

Docket No. 202306035 CLAIMS What is claimed is:

1. An object classification computer system comprising: a camera configured to capture an image of a physical environment, the image defining an object and noise, so as to define a naturally corrupted image; a memory having a plurality of application modules stored thereon; and a processor for executing the application modules, the application modules comprising: a noise classifier configured to determine one or more noise types of the noise in the naturally corrupted image; a plurality of generative adversarial networks (GANs) configured to denoise corrupted images; and a generative adversarial network (GAN) selector configured to, based on the one or more noise types, select one or more GANs from the plurality of GANs, so as to define one or more selected GANs.

2. The system as recited claim 1, wherein the one or more selected GANs are configured to denoise the naturally corrupted image, so as to generate respective denoised inputs.

3. The system as recited in claim 2, the application modules further comprising: a deep neural network (DNN) configured to generate an uncertainty for each of the respective denoised inputs, wherein the system is configured to classify the object in the naturally corrupted image, based on the denoised input associated with the uncertainty that is lower than the uncertainty of the other denoised inputs.

4. The system as recited in claim 1, wherein the noise is defined by snow, fog, rain, or glare.

5. The system as recited in claim 1, wherein the system is configured to train the GAN selector with first images that the plurality of GANs misclassified, so as to define misclassified data.Docket No. 202306035 6. The system as recited in claim 1, wherein the system is further configured to also train the GAN selector with second images that at least one of the plurality of GANs correctly classified, so as to define correctly classified data.

7. The system as recited in claim 6, wherein the first images define a specific object and a specific noise type, and the second images also define the specific object and the specific noise type.

8. The system as recited in claim 7, wherein the first images define a first noise intensity that is less than a second noise intensity defined by the second images, such that each of the first images are equivalent to respective second images except for the respective noise intensities.

9. A method performed by an object classification computer system that defines a plurality of generative adversarial networks (GANs) and a generative adversarial network (GAN) selector, the method comprising: capturing an image of a physical environment, the image defining an object and noise, so as to define a naturally corrupted image; determining one or more noise types of the noise in the naturally corrupted image; and the GAN selector, based on the one or more noise types, selecting one or more GANs from the plurality of GANs, so as to define one or more selected GANs.

10. The method as recited in claim 9, the method further comprising: the one or more selected GANs denoising the naturally corrupted image, so as to generate respective denoised inputs.

11. The method as recited in claim 10, the method further comprising: a deep neural network (DNN) generating an uncertainty for each of the respective denoised inputs; and classifying the object in the naturally corrupted image, based on the denoised input associated with the uncertainty that is lower than the uncertainty of the other denoised inputs.Docket No. 202306035 12. The method as recited in claim 9, wherein the noise is defined by snow, fog, rain, or glare.

13. The method as recited in claim 9, the method further comprising: training the GAN selector with first images that the plurality of GANs misclassified, so as to define misclassified data.

14. The method as recited in claim 13, the method further comprising: also training the GAN selector with second images that at least one of the plurality of GANs correctly classified, so as to define correctly classified data.

15. The method as recited in claim 14, wherein the first images define a first noise intensity that is less than a second noise intensity defined by the second images, such that each of the first images are equivalent to respective second images except for the respective noise intensities.

Images