Multi-modal large language model object illusion suppression method and related device

Abstract

The invention discloses a multi-modal large language model object hallucination suppression method and a related device. The method comprises the following steps: constructing a co-occurrence score matrix based on a rare strong correlation object pair score and a frequent stable object pair score; inputting the input image into the multi-modal large language model to obtain initial text description, and extracting a plurality of key objects; analyzing the key objects by applying the co-occurrence score matrix to obtain a high-risk illusion object set; and judging whether each high-risk illusion object exists in the input image or not, forming a visual evidence, further constructing an inhibitory cue word, and inputting the cue word into a model to obtain a final text description. A high-risk illusion object with high statistical relevance and weak visual evidence is accurately positioned through the co-occurrence score matrix, and excessive dependence of the model on a language statistical rule is reduced fundamentally. By means of the screening mechanism based on prior knowledge, a few high-risk targets are verified in a centralized mode through follow-up computing resources, and huge computing expenditure caused by indifference verification is overcome.

Description

Technical Field

[0001] This invention relates to the field of computer vision, and more particularly to a method and apparatus for suppressing object illusions in a multimodal large language model. Background Technology

[0002] Multimodal large language models have made significant progress in computer vision and natural language processing, and are widely used in fields such as intelligent healthcare, autonomous driving, and security monitoring. However, multimodal large language models commonly suffer from the object illusion problem, where the text descriptions generated by the model contain objects that do not match the actual content of the input image.

[0003] The main reasons for object illusions include two aspects: over-reliance on co-occurrence bias and insufficient visual information. Co-occurrence bias refers to the model learning strong associations between certain objects in the training data (such as "bench" and "chair," "person" and "backpack"), causing it to generate these associated objects even when direct visual evidence is lacking in the image. Figure 1 As shown, the model incorrectly generated "bench" and "backpack" in the initial description, even though they did not appear in the image. Insufficient visual information means that multimodal large language models compress and crop the original input image to a fixed resolution, resulting in the loss of a large amount of visual detail. This causes the model to rely more on linguistic statistical patterns than visual evidence when generating descriptions, thus producing object illusions.

[0004] Most existing solutions suppress object illusions during the decoding phase of the model output, failing to fundamentally reduce the over-reliance of multimodal large language models on co-occurrence bias and lacking sufficient attention to visual information. Some existing solutions improve visual perception by modifying the model architecture, such as using a higher-resolution visual encoder, but these solutions inevitably introduce huge computational overhead and memory consumption, and require model retraining, making them difficult to apply quickly to existing deployed models. Summary of the Invention

[0005] The purpose of this invention is to overcome the problems of the prior art and provide a method and related device for suppressing multimodal large language model object illusions.

[0006] The objective of this invention is achieved through the following technical solution: a method for suppressing object illusions in a multimodal large language model, the method comprising the following steps: A co-occurrence score matrix is ​​constructed based on the scores of rare strongly correlated object pairs and frequent stable object pairs. The scores of rare strongly correlated object pairs are calculated based on the deviation between the co-occurrence probability and statistical independence between objects, and the scores of frequent stable object pairs are calculated based on the overlap coefficient between objects. The input image is fed into a multimodal large language model to obtain an initial text description, and multiple key objects are extracted from the initial text description. Based on the co-occurrence score matrix analysis, key objects extracted from the initial text description are obtained to acquire a set of high-risk hallucination objects. Determine whether each high-risk hallucination object in the set of high-risk hallucination objects exists in the input image, and use the determination result to form visual evidence; Inhibitory cue words are constructed based on visual evidence and input into a multimodal large language model to obtain the final text description of the input image.

[0007] In one example, the key objects extracted from the initial text description based on the co-occurrence score matrix analysis to obtain a set of high-risk hallucination objects are replaced with: Locate the regions of each high-risk hallucination object in the input image and generate corresponding visual stimulus windows; Based on the center coordinates of each visual stimulus window, the key visual regions, including the visual stimulus windows, are reconstructed from the input image. Adaptive focus blurring is applied to the background portion of the key visual region to generate a focused image; Each focused image is input into a multimodal large language model and visual question answering is performed. All answers are then aggregated to form visual evidence.

[0008] In one example, the co-occurrence score matrix is ​​a linear weighted sum of scores for rare strongly correlated object pairs and scores for frequent stable object pairs. The construction of the co-occurrence score matrix also includes: Normalize the scores of rare strongly correlated object pairs and frequent stable object pairs respectively, and introduce penalty factors into the scores of rare strongly correlated object pairs and frequent stable object pairs respectively.

[0009] In one example, the rare strongly correlated object pair scores The calculation expression is: in, This represents the deviation between the co-occurrence probability and statistical independence of object pairs; yes Normalized values; express The penalty factor; Represents objects The occurrence count; Represents objects The occurrence count; Represents objects and objects The occurrence count; This indicates the total number of images.

[0010] In one example, the frequent stable object pairs fractions The calculation expression is: in, Represents objects , The overlap coefficient between them; express The penalty factor; Represents objects The occurrence count; Represents objects The occurrence count; Represents objects and objects The occurrence count.

[0011] In one example, key visual regions are aligned based on the input size of the multimodal large language model visual encoder.

[0012] In one example, the adaptive focus blurring of the background portion in the key visual region includes: Based on the Euclidean distance from each pixel within the key visual region to the center coordinates, the standard deviation of the adaptive Gaussian blur kernel corresponding to each pixel is calculated; the standard deviation is proportional to the Euclidean distance. An adaptive Gaussian blur kernel is generated based on the standard deviation of the adaptive Gaussian blur kernel. All background pixels in the key visual region are blurred based on the adaptive Gaussian blur kernel, while all foreground pixels in the visual region are preserved, generating the final focused image.

[0013] It should be further noted that the technical features corresponding to the above examples can be combined or replaced to form new technical solutions.

[0014] The present invention also includes a computer program product comprising a computer program that, when executed by a processor, implements the steps of the multimodal large language model object illusion suppression method formed by any or a combination of the above examples.

[0015] The present invention also includes a storage medium storing computer instructions that, when executed, perform the steps of the multimodal large language model object illusion suppression method formed by any or more of the above examples.

[0016] The present invention also includes a terminal comprising a memory and a processor, the memory storing computer instructions executable on the processor, wherein the processor, when executing the computer instructions, performs the steps of the multimodal large language model object illusion suppression method formed by any or more of the above examples.

[0017] Compared with the prior art, the beneficial effects of the present invention are: 1. Precise targeting and efficiency optimization: It has achieved a paradigm shift from "blind full verification" to "intelligent risk screening".

[0018] Specifically, the constructed co-occurrence score matrix comprehensively considers two key patterns: rare strong associations and frequent stable associations. By using the co-occurrence score matrix, it quantifies and actively intervenes in the inherent co-occurrence bias of multimodal large language models. This allows for the accurate identification of high-risk hallucinatory objects in the initial description that have strong statistical correlations but weak visual evidence, fundamentally reducing the model's over-reliance on linguistic statistical patterns. Furthermore, it constructs inhibitory cue words based on visual evidence to suppress hallucinations. Finally, this screening mechanism based on prior knowledge allows subsequent computational resources to be concentrated on verifying a few high-risk targets, overcoming the enormous computational overhead caused by indiscriminate verification of all generated content in existing technologies, and significantly improving computational efficiency.

[0019] 2. No training required and excellent compatibility: Provides a solution with zero training overhead and rapid deployment.

[0020] Specifically, the method of this invention is a post-processing framework that requires no fine-tuning, parameter updates, or architectural modifications to existing multimodal large language models. This not only completely avoids the enormous computational, time, and data collection costs associated with model retraining, but also eliminates the technical risks of knowledge loss or performance degradation that may result from adjusting model parameters. This method is widely compatible with various mainstream multimodal large language models, such as LLaVA-1.5 and MiniGPT-4, and can be quickly deployed in existing inference pipelines, demonstrating extremely high practical value and deployment flexibility.

[0021] 3. Visual information enhancement: Reliable visual evidence is reinforced at its source through high-resolution focused visual stimuli.

[0022] Specifically, by using high-resolution reconstruction and adaptive focusing blur processing, the intelligent visual information enhancement mechanism of the biological visual system, which "improves spatial resolution" and "transfers receptive fields", is simulated. This not only makes up for the loss of detail information in the multimodal large language model due to global image compression through local high-definition, but also suppresses misleading information that may exist in the background, greatly enhancing the visual signal intensity of the target object, thereby effectively cutting off the visual triggers that cause hallucinations.

[0023] 4. Introducing a penalty factor into the scores of rare strongly correlated object pairs can suppress spurious unexpected associations caused by extremely low co-occurrence rates, thereby improving the measurement accuracy of rare strongly correlated object pairs; introducing a penalty factor into the scores of frequent stable object pairs can reduce the overestimation of associations of relatively rare objects, thereby improving the measurement accuracy of frequent stable object pairs.

[0024] 5. Aligning key visual regions with the model's visual encoder input can improve the resolution of key visual regions without changing the architecture of the multimodal large language model, thereby further suppressing illusions. Attached Figure Description

[0025] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. The accompanying drawings are provided to provide a further understanding of the present application and constitute a part of the present application. The same reference numerals are used in these drawings to denote the same or similar parts. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation of the present application.

[0026] Figure 1 A schematic diagram illustrating object illusion correction in the generation of descriptions for multimodal large language models; Figure 2 A flowchart illustrating a method provided as an example of the present invention; Figure 3 A method flowchart is provided as a preferred example of the present invention; Figure 4 This is a comparison image of the high-resolution reconstruction before and after, provided as an example of the present invention; Figure 5 This is a comparison image showing the effect of adaptive focus blur before and after, as provided in an example of the present invention. Detailed Implementation

[0027] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] Furthermore, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0029] In one example, such as Figure 2 As shown, a method for suppressing object illusions in a multimodal large language model includes the following steps: S10: Construct a co-occurrence score matrix based on the scores of rare strongly correlated object pairs and frequent stable object pairs. The scores of rare strongly correlated object pairs are calculated based on the deviation between the co-occurrence probability and statistical independence between objects, and the scores of frequent stable object pairs are calculated based on the overlap coefficient between objects.

[0030] Specifically, the labeled object categories in the multimodal training dataset are extracted, and the co-occurrence frequency of each category is calculated to construct a basic co-occurrence frequency table. The data in the basic co-occurrence frequency table is smoothed, and the contribution strength of each object pair is calculated. The score for rare strongly correlated object pairs highlights uncommon but strongly correlated object pairs in the dataset; that is, the score for rare strongly correlated object pairs is highly sensitive to low-frequency but strongly correlated object pairs. The score for frequent and stable object pairs captures frequently and stably occurring object pairs in the dataset, and is more sensitive to stable and frequently coexisting object pairs. A co-occurrence score matrix is ​​constructed using these two complementary indicators to provide a more comprehensive model of object relationships.

[0031] S20: Input the input image into the multimodal large language model to obtain the initial text description, and extract multiple key objects from the initial text description.

[0032] Multimodal large language models typically expand the description of an input image by focusing on key objects in the input image. This example uses spacy to identify and extract the key objects mentioned in the generated description.

[0033] S30: Analyze the key objects extracted from the initial text description based on the co-occurrence score matrix to obtain a set of high-risk hallucination objects.

[0034] In this step, the co-occurrence score matrix between two key objects is used to determine whether it is greater than the adjustable hyperparameter. If it is, the two key objects are considered high-risk hallucination objects, thus obtaining a set of high-risk hallucination objects. For object pairs and If the co-occurrence score matrix ,but ,in It is an adjustable hyperparameter.

[0035] S40: Determine whether each high-risk hallucination object in the set of high-risk hallucination objects exists in the input image, and use the determination result to form visual evidence.

[0036] Optionally, a visual inspection model can be used to determine whether each high-risk hallucination object exists in the input image, thus obtaining visual evidence that conforms to the facts.

[0037] S50: Construct inhibitory cue words based on visual evidence and input them into a multimodal large language model to obtain the final text description of the input image.

[0038] In this step, visual evidence is used to correct the model output, that is, to prevent the multimodal large language model from mentioning the above-mentioned inhibitory cue words when describing the input image, thereby correcting the model output and obtaining the final text description of the input image.

[0039] In one example, the co-occurrence score matrix is ​​a linear weighted sum of scores for rare strongly correlated object pairs and scores for frequent stable object pairs. Constructing the co-occurrence score matrix also includes normalizing the scores for rare strongly correlated object pairs and frequent stable object pairs respectively, and introducing penalty factors into the scores for rare strongly correlated object pairs and frequent stable object pairs respectively. Then, the scores for rare strongly correlated object pairs are calculated. include: First, calculate the deviation between the co-occurrence probability and the statistical independence. : in, and Represents objects and The occurrence count; Represents objects and objects The occurrence count; This is the total number of images. Adding 1 smooths the image to avoid zero frequencies. Positive value. express and They occur simultaneously more frequently than expected when independent. To make the scores comparable across different frequencies, [the following is omitted]. Normalization process: In addition, a penalty factor was introduced. To suppress spurious unexpected associations caused by extremely low co-occurrence rates, at which point the object and Between Defined as: Calculate the fraction of frequent stable objects include: First, calculate the overlap coefficient between the two objects. To quantify the frequency of object pairs appearing together: in, It quantifies the proportion of infrequent objects appearing simultaneously with another object. However, when an object is relatively rare, it relies solely on... This may overestimate the correlation. To mitigate this, a penalty factor is also introduced here. Inspired by the sigmoid function, Designed to exert minimal influence on high-frequency objects (e.g., At the same time, a lower limit is set for low-frequency objects, at which point the penalty factor... for: Through merger ,make sure This reflects a more accurate and reliable measure of stable co-occurrence patterns. At this point, the co-occurrence score matrix is: in, Represents objects and The co-occurrence score between them, the higher the value, the stronger the association; It is the balance coefficient.

[0040] In one example, instead of using the judgment results to form visual evidence that each high-risk hallucination object in the high-risk hallucination object set exists in the input image, the following is replaced: S401: Locating a collection of high-risk hallucination objects For each high-risk hallucination object in the input image, a corresponding visual stimulus window is generated.

[0041] High-risk hallucinatory objects Using Grounding DINO to locate high-risk hallucinatory objects This generates a visual stimulus window. ,in and Indicates width and height. Indicates the center coordinates.

[0042] S402: Reconstruct key visual regions, including the visual stimulus windows, from the input image based on the center coordinates of each visual stimulus window.

[0043] Preferably, the key visual regions are aligned according to the input size of the multimodal large language model visual encoder, at which point a reconstructed visual region is obtained. Square key visual area The input image resolution of the multimodal large language model visual encoder (LLaVA-1.5 visual encoder) is... Input alignment is used to improve the resolution of key visual regions without changing the architecture of the multimodal large language model.

[0044] S403: Perform adaptive focus blurring on the background in the key visual region to generate a focused image.

[0045] Specifically, key visual areas Preserving the visual stimulus window The overall resolution is adjusted, while adaptive blur is applied to the remaining background to enhance high-risk hallucination objects. The stimulus signal is received and a focused image is generated. .

[0046] S404: Input each focused image into a multimodal large language model and perform visual question answering, then aggregate all answers to form visual evidence.

[0047] In obtaining a focused image Then, input into the multimodal large language model And one question: "Are there any high-risk hallucinatory objects in the image?" To obtain the answer, ask "?", then gather all the high-risk hallucination objects. The answers are compiled into a set of visual evidence. Finally, the visual evidence is used as basic facts to construct the prompt "Please describe this image in detail. Do not mention these objects:" ", used to obtain the final output from a multimodal large language model.

[0048] In one example, adaptive focus blurring is applied to the background portion of a key visual region, including: S4031: Calculate the standard deviation of the adaptive Gaussian blur kernel for each pixel based on the Euclidean distance from the center coordinates of each pixel within the key visual region; where the standard deviation of a pixel is larger the farther it is from the center, i.e., the standard deviation is proportional to the Euclidean distance. S4032: Generate an adaptive Gaussian blur kernel based on the standard deviation of the adaptive Gaussian blur kernel. Based on the adaptive Gaussian blur kernel, blur all background pixels in the key visual region and retain all foreground pixels in the visual region to generate the final focused image.

[0049] Specifically, in key visual areas Within the range, visual stimulus window While a small area might be occupied, a larger background can still induce object illusions, thus necessitating the introduction of adaptive focus blur. Unlike standard Gaussian blur, this invention modulates the blur intensity based on the distance of each pixel to the center. This adaptive suppression of the background guides multimodal large language models to focus more on the central region. Formally, given a size of... visual area and visual stimulation window Adaptive standard deviation for each pixel Defined as: in yes pixel coordinates in It is the center. These are hyperparameters. Then, a focused image is generated. : .

[0050] in, These are relative coordinates with respect to the center point; Indicates standard deviation The defined Gaussian blur kernel in relative coordinates The weight value at the location; and They are and median coordinate The pixel value at that location. Pixels closer to the center. ( Very small pixels will exhibit weak blurring; conversely, pixels farther from the center will be more blurred. This invention applies this adaptive blurring to gradually guide the attention of the multimodal large language model to the visual stimulus window. .

[0051] Combining the above examples, we obtain a preferred example of the present invention. This example uses the multimodal large language model LLaVA-1.5 as its foundation, running on a high-performance computing server equipped with an NVIDIA 3090 GPU. Each round of input to the multimodal large language model LLaVA-1.5 consists of an image and a prompt: "Please describe this image in detail." The original image is randomly sampled from the MSCOCO2014 validation set, such as... Figure 3 As shown, the object hallucination suppression method based on a multimodal large language model of co-occurrence-guided visual stimuli includes the following steps: S1: Dataset preprocessing and object statistics.

[0052] Preprocessing and object statistics were performed on the MSCOCO2014 validation set. Object categories were extracted from the labeled images, and the co-occurrence frequency of each category was calculated to construct a basic co-occurrence frequency table. In the first row of the basic co-occurrence frequency table {"category1": "person", "category2": "car", "count": 6062}, count represents the number of times "person" and "car" co-occur in the same image in the MSCOCO2014 validation set, which is 6062.

[0053] S2: Construct a co-occurrence score matrix based on the scores of rare strongly correlated object pairs and the scores of frequent stable object pairs.

[0054] Based on the fundamental co-occurrence frequency table, scores are obtained using rare strongly correlated object pairs. Frequent stable object pairs fractions These two complementary metrics calculate the co-occurrence intensity of each pair of objects, constructing a co-occurrence score matrix. In the MSCOCO2014 validation set, "keyboard" and "mouse" are uncommon but strongly correlated object pairs, while "person" and "vehicle" are frequent and stable co-occurrence pairs. Taking "bench" and "chair" from the MSCOCO2014 dataset as an example, their co-occurrence score matrix values ​​are calculated. The calculations show a rare strong association between "bench" and "chair". The score is 0.723, and the frequent stable score (FS) for "bench" and "chair" is 0.398. The balance coefficients in the co-occurrence score matrix are then analyzed. With the value set to 4, the co-occurrence score of "bench" and "chair" was calculated to be 0.561.

[0055] S3: Input the input image into the multimodal large language model to obtain the initial text description, and extract multiple key objects from the initial text description.

[0056] enter Figure 1 After adding the prompts, the initial description from the multimodal large language model is obtained: "This photo captures a group of boys enjoying a day on the beach, playing in the waves. Four boys are playing in the water, with the boy in the center standing on a surfboard in the middle of the scene. In the background, there are some benches and chairs, providing a place for people to relax and enjoy the scenery. A backpack can also be seen on the sand, possibly belonging to a beachgoer. In addition, there are houses further away."

[0057] The descriptions generated by the multimodal large language model were identified and extracted using spacy, and the key objects "surfboard, chair, bench, person, backpack, house" were extracted from the generated descriptions.

[0058] S4: Analyze the key objects extracted from the initial text description based on the co-occurrence score matrix to obtain a set of high-risk hallucination objects; For object pairs and If the co-occurrence score matrix ,but Query the co-occurrence score matrix in step S3. The key object pairs obtained have RS of 0.723 and FS of 0.398 for "bench" and "chair"; and RS of 0.519 and FS of 0.624 for "person" and "backpack". These two object pairs satisfy... Add to a collection of high-risk hallucination objects .

[0059] S5: Locate the regions of each high-risk hallucination object in the input image and generate corresponding visual stimulus windows.

[0060] For high-risk objects Using Grounding DINO for positioning This generates a visual stimulus window. ,in and Indicates width and height. The center coordinates are given as (279, 208), (450, 144), (517, 383), and (443, 219), respectively.

[0061] S6: Reconstruct key visual regions, including the visual stimulus windows, from the input image based on the center coordinates of each visual stimulus window.

[0062] High-risk hallucinatory objects Based on its center point A reconstructed image from the uncompressed original image Square key visual area Aligned with the visual encoder input of the multimodal large language model LLaVA-1.5. For example... Figure 4 The high-resolution reconstruction showed significant improvement, and the answers from the multimodal large language model were successfully corrected.

[0063] For the high-risk objects "bench", "chair", "person", and "backpack", four reconstructed images were created on the uncompressed original image, centered at (279, 208), (450, 144), (517, 383), and (443, 219) respectively. The key visual regions were optimized, ensuring that the resolution of the key visual regions was improved without changing the architecture of the multimodal large language model LLaVA-1.5.

[0064] S7: Adaptive focus blurring is applied to the background in the key visual region to generate a focused image.

[0065] For high-risk objects such as "bench", "chair", "person", and "backpack", based on the four corresponding key visual regions obtained in step S6, adaptive focus blur is introduced for each key visual region to obtain four focused images. This adaptive fuzziness is applied to gradually guide the attention of the multimodal large language model to the visual stimulus window of high-risk objects such as "bench," "chair," "person," and "backpack," as shown in the example. Figure 5 As shown, the adaptive focusing blurring process has a significant effect, and the answer from the multimodal large language model has been successfully corrected.

[0066] S8: Input each focused image into a multimodal large language model and perform visual question answering, then summarize all the answers to form visual evidence.

[0067] In obtaining focused images of high-risk objects such as "bench," "chair," "person," and "backpack." Then, first enter And one question: "Are there any high-risk hallucinatory objects in the image?" Given a multimodal large language model, obtain the answers "There is no bench in the image", "There is a chair in the image", "There is a person in the image", and "There is no backpack in the image", then collect all high-risk hallucination objects. The answers were compiled into a set of visual evidence.

[0068] S9: Construct inhibitory cue words based on visual evidence and input them into a multimodal large language model to obtain the final text description (corrected description) of the input image.

[0069] Visual evidence was used as basic facts to construct the cue "Please describe this image in detail. Do not mention these objects: bench, backpack," which was then used to obtain the final output from a multimodal large language model. The corrected multimodal large language model's response was: "The image shows a group of little boys having fun on the beach, with waves crashing against them. Four boys are playing in the water, and the boy in the middle is standing on a surfboard. Some chairs are placed in the back, providing a place to rest and enjoy the scenery. Houses can be seen further away, giving a lively and relaxed coastal atmosphere." The object illusion was corrected.

[0070] This invention extracts key objects of high-risk hallucinations by constructing a co-occurrence score matrix and enhances visual information by using high-resolution focused visual stimuli as visual evidence, which significantly improves the response ability of multimodal large language models. It also optimizes the object hallucination of multimodal large language models by jointly optimizing co-occurrence bias and visual information.

[0071] Experiments demonstrate that on a single NVIDIA 3090 GPU, when performing CHAIR evaluation on 500 images for LLaVA-1.5, the mainstream decoding intervention method OPERA has a total inference time of 19,310 seconds, while the method of this invention only requires 6,024 seconds, reducing the total time by 68.8% and achieving a significant efficiency improvement. This is mainly due to the precise targeting mechanism of the method of this invention, which eliminates the need for complex calculations for each generated token, thereby effectively suppressing hallucinations while minimizing unnecessary computational overhead.

[0072] In the CHAIR evaluation of the multimodal large language model LLaVA-1.5, the original results showed that its sentence-level CHAIR index (CS) was 50.2 and its instance-level CHAIR index (CI) was 14.9. After introducing the method of this invention, both indices were significantly optimized, with the sentence-level index (CS) decreasing to 44.2 and the instance-level index (CI) decreasing to 12.8. Calculations show that this invention improves LLaVA-1.5's sentence-level index by 12.0% and its instance-level index by 14.1% in CHAIR evaluation. Comparative experiments were conducted with popular methods such as OPERA, HALC, SID, LURE, and WoodPecker, and the evaluation results of the sentence-level CHAIR index (CS) and instance-level CHAIR index (CI) are shown in Tables 1-2. Table 1. Results of the LLaVA-1.5 model and its mainstream applications on the CHAIR index. Table 2. Results of the MiniGPT-4 model and its mainstream applications on the CHAIR index. Based on the experimental data in Tables 1 and 2, the experimental results of this invention are superior to the currently popular methods OPERA, HALC, SID, LURE, and WoodPecker, indicating that this invention has a significant effect in reducing large model illusions and improving description accuracy.

[0073] The present invention also provides a computer program product, comprising a computer program that, when executed by a processor, implements the steps of the multimodal large language model object illusion suppression method formed by any or a combination of the above examples. The processor may be a single-core or multi-core central processing unit or a specific integrated circuit, or one or more integrated circuits configured to implement the present invention.

[0074] The present invention also provides a storage medium having the same inventive concept as the multimodal large language model object illusion suppression method formed by any or more of the above examples, wherein computer instructions are stored thereon, which, when executed, perform the steps of the multimodal large language model object illusion suppression method formed by any or more of the above examples.

[0075] Based on this understanding, the technical solution of this embodiment, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0076] This invention also provides a terminal having the same inventive concept as any or a combination of examples corresponding to the above-described method for suppressing object illusions in a multimodal large language model, including a memory and a processor. The memory stores computer instructions executable on the processor, and the processor executes the steps of the above-described method for suppressing object illusions in a multimodal large language model when executing the computer instructions. The processor may be a single-core or multi-core central processing unit or a specific integrated circuit, or one or more integrated circuits configured to implement this invention.

[0077] In one example, the terminal, i.e., the electronic device, is represented in the form of a general-purpose computing device. The components of the electronic device may include, but are not limited to: at least one processing unit (processor) mentioned above, at least one storage unit mentioned above, and a bus connecting different system components (including storage units and processing units).

[0078] The storage unit stores program code that can be executed by the processing unit, causing the processing unit to perform the steps described in the "Exemplary Methods" section of this specification according to various exemplary embodiments of the present invention. For example, the processing unit can perform the above-described method for suppressing object illusions in a multimodal large language model.

[0079] The storage unit may include a readable medium in the form of a volatile storage unit, such as a random access memory (RAM) 3201 and / or a cache storage unit, and may further include a read-only memory (ROM).

[0080] The storage unit may also include a program / utility having a set (at least one) of program modules, including but not limited to: an operating system, one or more applications, other program modules, and program data, each or some combination of these examples may include an implementation of a network environment.

[0081] A bus can represent one or more of several types of bus structures, including a memory cell bus or memory cell controller, a peripheral bus, a graphics acceleration port, a processing unit, or a local bus that uses any of the various bus structures.

[0082] The electronic device can also communicate with one or more external devices (such as keyboards, pointing devices, Bluetooth devices, etc.), one or more devices that enable a user to interact with the electronic device, and / or any device that enables the electronic device to communicate with one or more other computing devices (such as routers, modems, etc.). This communication can be performed via input / output (I / O) interfaces. Furthermore, the electronic device can communicate with one or more networks (such as local area networks (LANs), wide area networks (WANs), and / or public networks, such as the Internet) via a network adapter. The network adapter communicates with other modules of the electronic device via a bus. It should be understood that other hardware and / or software modules can be used in conjunction with the electronic device, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0083] Through the above description, those skilled in the art will readily understand that the exemplary embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solution according to this exemplary embodiment can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) or on a network, including several instructions to cause a computing device (such as a personal computer, server, terminal device, or network device, etc.) to execute the method of the exemplary embodiment of this application.

[0084] The above detailed embodiments are a description of the present invention. It should not be considered that the specific embodiments of the present invention are limited to these descriptions. For those skilled in the art, several simple deductions and substitutions can be made without departing from the concept of the present invention, and all of these should be considered to fall within the protection scope of the present invention.

Claims

1. A method for suppressing object hallucinations in a multimodal large language model, characterized in that, Includes the following steps: A co-occurrence score matrix is ​​constructed based on the scores of rare strongly correlated object pairs and frequent stable object pairs. The scores of rare strongly correlated object pairs are calculated based on the deviation between the co-occurrence probability and statistical independence between objects, and the scores of frequent stable object pairs are calculated based on the overlap coefficient between objects. The input image is fed into a multimodal large language model to obtain an initial text description, and multiple key objects are extracted from the initial text description. Based on the co-occurrence score matrix analysis, key objects extracted from the initial text description are obtained to acquire a set of high-risk hallucination objects. Determine whether each high-risk hallucination object in the set of high-risk hallucination objects exists in the input image, and use the determination result to form visual evidence; Inhibitory cue words are constructed based on visual evidence and input into a multimodal large language model to obtain the final text description of the input image.

2. The method for suppressing object hallucinations in a multimodal large language model according to claim 1, characterized in that, The key objects extracted from the initial text description based on the co-occurrence score matrix analysis to obtain the set of high-risk hallucination objects are replaced with: Locate the regions of each high-risk hallucination object in the input image and generate corresponding visual stimulus windows; Based on the center coordinates of each visual stimulus window, the key visual regions, including the visual stimulus windows, are reconstructed from the input image. Adaptive focus blurring is applied to the background portion of the key visual region to generate a focused image; Each focused image is input into a multimodal large language model and visual question answering is performed. All answers are then aggregated to form visual evidence.

3. The method for suppressing object hallucinations in a multimodal large language model according to claim 1 or 2, characterized in that, The co-occurrence score matrix is ​​a linear weighted sum of scores for rare strongly correlated object pairs and scores for frequent stable object pairs. The construction of the co-occurrence score matrix also includes: Normalize the scores of rare strongly correlated object pairs and frequent stable object pairs respectively, and introduce penalty factors into the scores of rare strongly correlated object pairs and frequent stable object pairs respectively.

4. The method for suppressing object hallucinations in a multimodal large language model according to claim 3, characterized in that, The rare strongly correlated object pairs score The calculation expression is: in, This represents the deviation between the co-occurrence probability and statistical independence of object pairs; yes Normalized values; express The penalty factor; Represents objects The occurrence count; Represents objects The occurrence count; Represents objects and objects The occurrence count; This indicates the total number of images.

5. The method for suppressing object hallucinations in a multimodal large language model according to claim 3, characterized in that, The frequent stable object pairs fraction The calculation expression is: in, Represents objects , The overlap coefficient between them; express The penalty factor; Represents objects The occurrence count; Represents objects The occurrence count; Represents objects and objects The occurrence count.

6. The method for suppressing object hallucinations in a multimodal large language model according to claim 2, characterized in that, The key visual regions are aligned based on the input size of the multimodal large language model visual encoder.

7. The method for suppressing object hallucinations in a multimodal large language model according to claim 2, characterized in that, The adaptive focus blurring process for the background portion in the key visual region includes: Based on the Euclidean distance from each pixel within the key visual region to the center coordinates, the standard deviation of the adaptive Gaussian blur kernel corresponding to each pixel is calculated; the standard deviation is proportional to the Euclidean distance. An adaptive Gaussian blur kernel is generated based on the standard deviation of the adaptive Gaussian blur kernel. All background pixels in the key visual region are blurred based on the adaptive Gaussian blur kernel, while all foreground pixels in the visual region are preserved, generating the final focused image.

8. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the multimodal large language model object illusion suppression method according to any one of claims 1-7.

9. A storage medium storing computer instructions thereon, characterized in that, When the computer instructions are executed, they perform the steps of the multimodal large language model object illusion suppression method according to any one of claims 1-7.

10. A terminal comprising a memory and a processor, wherein the memory stores computer instructions executable on the processor, characterized in that, When the processor executes the computer instructions, it performs the steps of the multimodal large language model object illusion suppression method according to any one of claims 1-7.

Images