An oligogenic chicken phenotype feature detection and identification method based on a multi-modal large model

Abstract

The application discloses a kind of oligogenic chicken phenotype feature detection and identification method based on multimodal large model, it is related to computer vision and artificial intelligence technical field, including: obtaining the image data and day age of cage chicken to be identified in henhouse;With the input of the pre-training oligogenic chicken accurate identification multimodal large model, the image data and day age of cage chicken to be identified and the related problem text of oligogenic chicken identification, obtain the identification result and coordinate of oligogenic chicken.The application realizes the efficient, high-precision oligogenic chicken identification in large-scale henhouse, provides a theoretical basis for realizing the oligogenic chicken accurate identification function of large-scale henhouse inspection robot, helps to promote the wide application of inspection robot in large-scale henhouse, provides key technical support for egg poultry intelligent breeding.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer vision and artificial intelligence, and in particular to a kind of oligogenic chicken phenotype feature detection and identification method based on multi-modal large model. BACKGROUND

[0002] In the process of large-scale poultry breeding, the detection and selection of oligogenic chickens (which have a much lower egg production rate than normal during the egg-laying period) cannot be ignored. According to statistics, oligogenic chickens account for about 3% of large-scale chicken flocks, severely affecting breeding efficiency and industry economic benefits. According to data such as the number of laying hens in 2024, average egg prices, average feed consumption, water consumption, and carbon emissions, "eating only feed and not laying eggs" oligogenic chickens consume an additional 230,000 tons of feed and 214 million tons of water per year, emit an additional 380,000 tons of CO2, and result in direct economic losses of nearly 10 billion yuan.

[0003] Currently, more than 80% of laying hens in China are raised in intensive cage systems, and some large-scale farms use the "one touch five look" method, especially the "night touch chicken" method, to identify oligogenic chickens in the flock and promptly eliminate them to improve breeding efficiency. However, due to the high stress on the chickens, low identification efficiency, and high labor intensity, this method cannot be widely applied. Therefore, it is necessary to use intelligent means to detect oligogenic chickens to reduce economic losses.

[0004] Some existing intelligent detection methods for oligogenic chickens rely solely on a single phenotype (such as comb color and size) for judgment. Such detection methods face two major limitations in practical application scenarios: first, detection models based on a single phenotype feature are highly sensitive and easily affected by environmental factors; second, in intensive cage systems, chickens often block each other, and once the key detection features are blocked, the model is prone to misjudgment. In addition, existing detection technologies mainly analyze the phenotype of the chicken, but have not effectively analyzed the non-linear relationship between the chicken's phenotype features and age, and still rely on human experience to determine the best identification time, which to some extent limits the development of end-to-end models and restricts the application of oligogenic chicken identification functions in actual chicken coops. SUMMARY

[0005] To address the above deficiencies in the prior art, the present application provides a kind of oligogenic chicken phenotype feature detection and identification method based on multi-modal large model, which solves the problem of low recognition accuracy in the prior art.

[0006] To achieve the above-mentioned purposes, the technical solution adopted by the present application is as follows: a kind of oligogenic chicken phenotype feature detection and identification method based on multi-modal large model, comprising:

[0007] obtaining image data and age of the cage chickens to be identified in the chicken coop;

[0008] The relevant question text for identifying low-producing chickens, the image data of the caged chickens to be identified, and their age are used as inputs to a pre-trained multimodal large model for accurate identification of low-producing chickens, so as to obtain the identification results and coordinates of low-producing chickens.

[0009] The beneficial effects of this invention are as follows:

[0010] This study employs a multimodal large-scale model for visual-text joint analysis of phenotypic features in low-producing chickens. By leveraging textual information to enhance the visual feature signals of low-producing chickens, it effectively avoids the problems of fuzzy features and inaccurate analysis caused by single-modal recognition. Age information is introduced to analyze the nonlinear changes in phenotypic features of low-producing chickens at different physiological stages. Egg production rate is incorporated as an auxiliary signal during inference to enhance the model's sensitivity. Ultimately, this achieves efficient and high-precision identification of low-producing chickens in large-scale chicken houses. This provides a theoretical basis for realizing the accurate identification function of low-producing chickens in large-scale chicken house inspection robots, and helps promote the widespread application of inspection robots in large-scale chicken houses, providing key technical support for intelligent poultry farming. Attached Figure Description

[0011] Figure 1 A flowchart of a method for detecting and identifying phenotypic features of oligoprolific chickens based on a multimodal large model is provided for an embodiment.

[0012] Figure 2 This is a schematic diagram of the working of the multimodal large model for accurate identification of low-yield chickens provided in the embodiment;

[0013] Figure 3 A schematic diagram illustrating the labeling and segmentation of phenotypic regions in low-yield chickens provided in this embodiment;

[0014] Figure 4 Attention heatmap for identifying oligoprolific chickens based on phenotypic features, provided as an example. Detailed Implementation

[0015] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0016] like Figure 1 As shown, in one embodiment of the present invention, a method for detecting and identifying oligoprolific chicken phenotypic features based on a multimodal large model includes the following steps:

[0017] S1. Construct a large-scale multimodal model for accurate identification of low-producing chickens:

[0018] like Figure 2As shown, the oligoparent chicken accurate identification multi-modal large model includes a visual encoder (the visual encoder is a double-branch structure, including a VGGT and a CLIP image encoder), a projection layer, and a base large language model (the base large language model can be a GPT-4, Llama, Qwen 2.5, etc. mainstream architecture, so that the oligoparent chicken multi-modal large model can be upgraded and iterated synchronously with the upgrade of the base model, maintaining its high performance). The day-old information projection layer, the depth projection layer, and the visual text projection layer are all projection layers, wherein the day-old information projection layer is a linear layer structure, and the depth projection layer and the visual text projection layer are both two-layer MLP structures.

[0019] S2, acquiring image data of cage-bred chickens in a chicken coop, generating description image text pair data based on a general large model, and converting the image text pair into a high-quality instruction following data set:

[0020] Acquiring image data of cage-bred chickens in a chicken coop; using a general large model to generate description image text pair data based on the acquired image data, and converting the image text pair data into a high-quality instruction following data set. When constructing the high-quality instruction following data set, initial description image text pair data is generated based on ChatGPT and Qwen 2.5 VL, specifically including: using image titles, bounding boxes, day-old information of oligoparent chickens, and coordinates of key phenotype parts (such as Figure 3 、 Figure 4 As shown, the comb, beak, eye, body area, and chicken claw can be used as phenotype parts for identifying oligoparent chickens) as symbolic representations, and the image is encoded into a sequence recognizable by a large language model. Three types of instruction following data, dialogue, detailed description, and complex reasoning, are designed, and the description is fine-tuned by experienced veterinarians according to the format to construct a high-quality image-text pair data set, which requires clearly describing the characteristics of different phenotype parts of each chicken and being able to make a judgment on whether it is an oligoparent chicken based on these phenotype characteristics. Query the general large model through context learning to generate high-quality instruction following data.

[0021] Context type 1 (image title):

[0022] There are eight chickens in a cage, they are full of energy, lively in appearance, and look healthy.

[0023] There are six eggs in the egg-laying area in front of the chicken coop.

[0024] In a chicken coop, a group of chickens are blocking each other, and three of them are drinking water.

[0025] There is a chicken lying down in the chicken group, with an enlarged abdomen and dull eyes, suspected to be an oligoparent chicken.

[0026] The chickens in the chicken coop are competing for feed, and some of them have partially lost their feathers.

[0027] Context Type 2 (Supervised Signal Coordinates):

[0028] Day of age {440 days}, target frame for low-producing chicken [0.681, 0.242, 0.774, 0.694], top of comb for low-producing chicken [0.758, 0.413, 0.845, 0.69], eyes for low-producing chicken [0.63, 0.222, 0.686, 0.516], beak tip for low-producing chicken [0.755, 0.413, 0.846, 0.692], moustache for low-producing chicken [0.384, 0.696, 0.485, 0.914], shank for low-producing chicken [0.1, 0.497, 0.173, 0.579], egg laying pen [0.261, 0.101, 0.787, 0.626].

[0029] Generated instruction following dataset includes three types: dialogue, detailed description, and complex reasoning, for example:

[0030] Type 1: Dialogue

[0031] Question: Please describe the scene in the picture.

[0032] Answer: The picture shows a chicken coop with limited interior space, slightly dark lighting, and several chickens moving inside, which may be inside a productive chicken house.

[0033] Question: How many chickens are in the picture?

[0034] Answer: There are eight or nine chickens in the picture, four of which are near the feeding trough, two of which are active in the middle area, and two or three of which are in the dark corner of the coop, one of which is severely obscured.

[0035] Type 2: Detailed Description

[0036] There are eight brown egg-laying hens in the wire chicken coop. Three chickens on the left side are fighting for food in the feeding trough, with bright red erect combs; two chickens in the middle have soft abdomens and disheveled feathers standing upright; a chicken in the corner on the right side is curled up: the comb is pale and atrophied, the beak and claws are still yellow, the feathers are close to the body, the head is low, the neck is curled, and the eyes are closed.

[0037] There are eight egg-laying hens in the chicken coop. Two chickens in the front are lowering their heads to peck, with bright red combs; four chickens in the middle are active, with full abdomens; a chicken in the back is standing close to the wall with drooping wings, with a bright yellow beak, a white comb, and disheveled feathers.

[0038] Type 3: Complex Reasoning

[0039] Question: Please tell me if there are any low-producing chickens in this image. If so, output their coordinates and the basis for your judgment.

[0040] Answer: Seven roosters in the flock have red combs, white legs and feet, and wide abdomens; one rooster is crouched in the corner of the cage, with a small and pale comb, bright yellow legs and feet, and a narrow and hard abdomen, with no feather loss. The rooster in the corner has a pale and atrophic comb, a yellowish face, and still yellowish legs and feet, with smooth and shiny feathers, which is likely to be a low-producing hen, with coordinates [0.660, 0.420, 0.486, 0.615]. Since the current age is close to 400 days, the nutrients ingested by the low-producing hen after not laying eggs are mainly used for maintaining its own metabolism, so it shows the phenomenon of smooth and shiny feathers. At the same time, the yellow pigment ingested in the feed cannot be transferred to the egg yolk, and thus is deposited in the beak, legs and feet, etc., showing a distinct bright yellow color.

[0041] S3, training a low-producing hen precise identification multi-modal large model according to instruction following data sets;

[0042] The training of the low-producing hen precise identification multi-modal large model includes a first stage and a second stage; in the first stage, the visual encoder and the base large language model are frozen, and the projection layer is trained using single-round instruction following data (such as "please describe the image" + artificial annotation description), so that the visual features and the text features are expressed in the same semantic space;

[0043] In the second stage, the visual encoder is frozen, and the projection layer and the base large language model are trained using the generated multi-round instruction following data (including description, reasoning, dialogue, etc.), and the weights of the projection layer and the base large language model are updated;

[0044] The specific training process of the second stage is as follows:

[0045] The generated high-quality instruction data set is used as the input of the low-producing hen precise identification multi-modal large model, and its expression is as follows:

[0046]

[0047] wherein, is the input sequence, is the instruction text, is the target answer;

[0048] The hidden state is generated as follows:

[0049]

[0050] wherein, represents the processing process of the structure;

[0051] The token prediction is performed according to the hidden state, and its expression is as follows:

[0052] ​

[0053] where, denotes the probability of predicting the next token under given conditions , denotes the token of the t th step of regression, denotes all tokens before the t th step of regression; denotes the softmax function, denotes the output layer weight of the base large language model;

[0054] The likelihood probability of the target answer sequence is maximized according to the token prediction probability, and its expression is:

[0055]

[0056] where, denotes the likelihood probability of the target answer sequence, denotes the total number of tokens of the target answer, denotes the natural logarithm, denotes all parameters of the projection layer and the base large language model, denotes the i th token of the target answer, i denotes the token index.

[0057] S4, the relevant problem text of the few-production chicken, the image data of the to-be-identified cage chicken and its age are taken as the input of the pre-trained few-production chicken accurate identification multi-modal large model to obtain the identification result and coordinates of the few-production chicken;

[0058] Specifically: a. The image data of the to-be-identified cage chicken in the chicken house is collected by the inspection robot, and the image data of the to-be-identified cage chicken is input into the visual encoder;

[0059] b. The image data is processed by VGGT to predict the depth information, and the output feature map of the DPT head in VGGT is extracted as the depth feature , and its expression is:

[0060]

[0061] where, denotes the processing process of the image data by VGGT, denotes a feature map with a height of , a width of , and a channel number of ; and

[0062] c. The RGB image feature of the image data is obtained by the CLIP image encoder​ , where

[0063]

[0064] wherein, represents the processing procedure of the CLIP image encoder on the image data; represents the number of patches of the image data, represents the original dimension of the CLIP feature, represents a matrix feature map with rows and columns;

[0065] d. mapping the depth feature to the visual feature space through a depth projection layer to obtain a mapped feature , where

[0066]

[0067] wherein, , respectively represent the weight matrix of the first layer MLP structure and the second layer MLP structure in the depth projection layer; represents an activation function; , respectively represent the bias vector of the first layer MLP structure and the second layer MLP structure in the depth projection layer; represents the number of tokens of the depth feature; represents a feature matrix with rows and columns;

[0068] e. aligning and channel splicing the mapped feature with the RGB image feature to obtain a visual feature containing depth information , where

[0069]

[0070] wherein, represents feature fusion; represents a feature matrix with rows and columns;

[0071] f. discretely grouping the physiological stages of the laying hens to obtain day-old grouping of 5 different stages;

[0072]

[0073] wherein, for grouping index, for the minimum age value in the data set, for the age bin interval, representing the age metadata, representing the data set is discretized into groups.

[0074] The age bin can allow partial overlap within each group to alleviate the impact of data abnormal distribution on model overfitting during training, for example:

[0075] Discretized groups: {AGE_BINS=[(16,19),(20,28),(29,43),(44,55),(56,100)]};

[0076] Groups with overlap:

[0077] {AGE_BINS_OVERLAP=[(16,19),(18,28),(27,43),(42,55),(54,100)]}.

[0078] g. Token embedding of age groups at different stages through the age information projection layer, obtaining age features aligned with visual features , whose expression is:

[0079] ;

[0080] where, represents the embedding function, represents the age metadata, represents the obtained age groups at different stages; represents a feature matrix with 1 row and columns;

[0081] h. Channel splicing of age features and visual features , obtaining visual features with spliced age information, so that the projected visual features can be "understood" by the language model, whose expression is:

[0082]

[0083] where, represents a feature matrix with rows and columns;

[0084] i. Token embedding of text description data of the corresponding question, obtaining text features ;

[0085] j. Concatenate the visual feature with the day-old information Input the visual text projection layer and the text feature for dimension alignment to obtain the feature , the expression of which is:

[0086] ;

[0087] , wherein, , respectively represent the weight matrix of the first layer MLP structure and the second layer MLP structure in the visual text projection layer; , respectively represent the bias vector of the first layer MLP structure and the second layer MLP structure in the visual text projection layer; is the unified feature dimension of the multi-modal large model for accurate identification of oligopregnant hens; represents a feature matrix with rows and columns;

[0088] k. Concatenate the feature with the text feature to obtain the cross-modal sequence information , the expression of which is:

[0089]

[0090] , wherein, represents the length of the text sequence, represents a feature matrix with rows and columns.

[0091] l. Take the aligned cross-modal sequence information as the input of the base large language model to obtain the identification result and coordinates of the oligopregnant hen.

[0092] The input of the multi-modal large model for accurate identification of oligopregnant hens is the image data of the cage hen to be identified, the day-old information, and the corresponding problem description. For example, the visual feature, the day-old metadata, and the text feature (such as the question "describe this picture") are concatenated to form a joint input sequence: {[IMG_1, IMG_2,..., IMG_256] + [age]+ [Q1, Q2,...,Qn]}.

[0093] When the multi-modal large model for accurate identification of oligopregnant hens identifies oligopregnant hens, a sensitivity coefficient of the egg-laying rate of each cage is introduced to dynamically adjust the identification threshold of oligopregnant hens in different cages; when the single-cage egg-laying rate is lower than the average egg-laying rate, the probability of the existence of oligopregnant hens is increased, and the sensitivity coefficient of the corresponding cage is increased, otherwise the sensitivity coefficient of the corresponding cage is decreased.

[0094] Adaptively adjust the threshold of identifying oligogenic chickens by introducing temperature parameter scaling original logits or dynamic text prompts;

[0095] The logits probability distribution scaled by temperature coefficient T is:

[0096]

[0097]

[0098]

[0099]

[0100] wherein, is the logits probability distribution scaled by temperature coefficient T, is a natural constant, represents logits scaled by temperature coefficient T, represents the number of logits, represents original logits, represents temperature parameter, represents sensitivity adjustment coefficient, represents egg laying rate bias, represents current cage egg laying rate, represents average egg laying rate.

[0101] Dynamic text prompts are, for example:

[0102] Sensitivity prompts are:

[0103] {Single cage egg laying rate ≤ 50%: sensitivity = "Identify any possible oligogenic chicken, even if not very sure"}

[0104] {50% ≤ single cage egg laying rate ≤ 70%: sensitivity = "Identify individuals who look like oligogenic chickens"}

[0105] {Single cage egg laying rate ≤ average egg laying rate: sensitivity = "Identify obvious oligogenic chickens"}

[0106] {Others: sensitivity = "Only identify very sure oligogenic chickens"}

[0107] Prompt prompt words: "[AGE_{age}] The egg laying rate of this cage is lower than the average value {}%, please {sensitivity}".

[0108] The multimodal large model related in the application can be deployed into an inspection robot device, and through setting an inspection task, efficient inspection of the few laying hens can be realized all day. Meanwhile, the model has strong expansibility, and in the future, only by adding the image and text data of dead hens and weak hens, the dead hen and weak hen identification function of the model can be expanded, and the common abnormal hen data is unified into the multimodal large model framework, and the vertical multimodal large model special for poultry breeding is fine-tuned.

[0109] In summary, the application realizes efficient and high-precision identification of few laying hens in a large-scale henhouse, provides a theoretical basis for realizing the precise identification function of few laying hens of a large-scale henhouse inspection robot, helps to promote the wide application of the inspection robot in the large-scale henhouse, and provides key technical support for egg and poultry intelligent breeding.

Claims

1. A method for detecting and identifying phenotypic features of oligoprolific chickens based on a multimodal large model, characterized in that, include: Acquire image data and age of caged chickens to be identified in the chicken coop; The relevant question text for identifying low-producing chickens and the image data of caged chickens to be identified are used as inputs to a pre-trained multimodal large model for accurate identification of low-producing chickens. Age information is introduced to analyze the nonlinear changes in phenotypic features of low-producing chickens at different physiological stages, and the identification results and coordinates of low-producing chickens are obtained. The large-scale multimodal model for accurate identification of low-yield chickens includes a visual encoder, a projection layer, and a base large language model connected in sequence. The visual encoder has a dual-branch structure, with one branch being a VGGT and the other a CLIP image encoder. The input of the visual encoder serves as the input of the VGGT and CLIP image encoders, and the outputs of the VGGT and CLIP image encoders serve as the output of the visual encoder. The age information projection layer, the depth projection layer, and the visual text projection layer are all projection layers, with the age information projection layer being a single-layer MLP structure, and the depth projection layer and the visual text projection layer being two-layer MLP structures. The age information projection layer is used to group the age at different stages and embed tokens to obtain age features aligned with visual features; The age-based groupings for different stages are obtained by discretizing the physiological stages of laying hens, allowing for partial overlap within each group. The grouping expression is as follows: in, For grouped indexes, The smallest age value in the dataset. Grouping by age interval, Represents age metadata, Indicates the dataset Discretize and group; When the multimodal large model for accurate identification of oligo-producing chickens is used to identify oligo-producing chickens, the sensitivity coefficient of the egg production rate of each cage is introduced to dynamically adjust the judgment threshold for oligo-producing chickens in different cages. When the egg production rate of a single cage is lower than the average egg production rate, the probability of the presence of oligo-producing chickens increases, and the sensitivity coefficient of the corresponding cage is increased; otherwise, the sensitivity coefficient of the corresponding cage is decreased. The threshold for determining oligo-producing chickens can be adaptively adjusted by introducing temperature parameters to scale the original logits or by using dynamic text prompts. The logits probability distribution after scaling by the temperature coefficient T is: in, This is the logits probability distribution scaled by the temperature coefficient T. It is a natural constant. This represents logits scaled by the temperature coefficient T. Indicates the number of logits. Represents the original logits. Indicates temperature parameter, This represents the sensitivity adjustment coefficient. This indicates a deviation in egg production rate. This indicates the current egg production rate in the chicken cage. This indicates the average egg production rate.

2. The method according to claim 1, characterized in that, The specific method for obtaining the identification results and coordinates of oligo-producing chickens is as follows: the image and text feature spaces are gradually aligned using a projection layer to obtain cross-modal sequence information; the aligned cross-modal sequence information is used as the input of the base large language model to obtain the identification results and coordinates of oligo-producing chickens.

3. The method according to claim 2, characterized in that, The specific method for progressively aligning the image and text feature spaces using projection layers is as follows: Image data of caged chickens to be identified Input visual encoder; The image data was used to predict depth information using VGGT, and the output feature map of the DPT head in VGGT was extracted as the depth feature. Its expression is: in, This indicates the VGGT process for processing image data. Indicates height as Width is The number of channels is Feature map; The RGB image features of the image data are obtained using the CLIP image encoder. Its expression is: in, This indicates the process by which the CLIP image encoder processes image data; This indicates the number of image data patches. Represents the original dimension of the CLIP feature. Indicates the number of rows. The number of columns is Matrix feature map; The depth features are mapped to the visual feature space through a depth projection layer to obtain the mapped features. Its expression is: in, , These represent the weight matrices of the first and second MLP structures in the depth projection layer, respectively. express Activation function; , These represent the bias vectors of the first and second MLP structures in the depth projection layer, respectively. The number of tokens representing deep features; Indicates the number of rows. The number of columns is The feature matrix; Mapping features With RGB image features Align and stitch the channels together to obtain visual features containing depth information. Its expression is: in, Indicates feature fusion; Indicates the number of rows. The number of columns is The feature matrix; The physiological stages of laying hens were discretized and grouped to obtain age-grouped groups for different stages; By using a projection layer to group ages at different stages and embedding tokens, we obtain visual features. Aligned age characteristics Its expression is: ； in, Indicates an embedded function. Represents age metadata, This indicates the different age groups obtained at different stages; This indicates that the number of rows is 1 and the number of columns is... The feature matrix; Age characteristics With visual features By performing channel stitching, visual features that incorporate age information are obtained. Its expression is: in, Indicates the number of rows. The number of columns is The feature matrix; Text features are obtained by embedding tokens into the text description data. ; Visual features that combine age information The input visual text projection layer is dimensionally aligned with the text features to obtain the features. Its expression is: ； in, , These represent the weight matrices of the first and second MLP structures in the visual text projection layer, respectively. , These represent the bias vectors of the first and second MLP structures in the visual text projection layer, respectively. To accurately identify the unified feature dimensions of a multimodal large model for low-yield chickens; Indicates the number of rows. The number of columns is The feature matrix; Features Text features Channel splicing is performed to obtain cross-modal sequence information. Its expression is: in, Indicates the length of the text sequence. Indicates the number of rows. The number of columns is The characteristic matrix of .

4. The method according to claim 3, characterized in that, The specific training process for the multimodal large model for accurate identification of oligo-producing chickens is as follows: Acquire image data of caged chickens in chicken coops; use a general large model to generate image-text pairs based on the acquired image data, and convert the image-text pairs into a command-following dataset; input the command-following dataset into a multimodal large model for accurate identification of low-producing chickens for training; The process of generating descriptive image text pairs from the acquired image data using a general large model involves: using the image data's title, bounding box, and coordinates of the age of low-producing chickens and key phenotypic regions as symbolic representations to encode the acquired image data into a sequence recognizable by the general large language model; querying the general large model through context learning based on the encoded sequence to generate a command-following dataset, which includes three types: dialogue, detailed description, and complex reasoning; and having experienced farmers adjust the generated command-following dataset to obtain the final command-following dataset.

5. The method according to claim 4, characterized in that, The training of the multimodal large model for accurate identification of oligo-producing chickens includes a first stage and a second stage. In the first stage, the visual encoder and the base large language model are frozen, and training is carried out using single-round instruction following data and only the projection layer is trained, so that visual features and text features are expressed in the same semantic space. The second stage freezes the visual encoder and uses the generated multi-round instruction following data to perform autoregressive training on the projection layer and the pedestal large language model, while updating the weights of the projection layer and the pedestal large language model. The specific training process for the second phase is as follows: The generated instruction dataset is used as input to the multimodal large model for accurate identification of oligo-producing chickens, and its expression is: in, Given the input sequence, This is the instruction text. For the target answer; Generate hidden state : in, express The process of structural processing; The expression for token prediction based on the hidden state is: in, Indicates under given conditions Predict the probability of the next token. Indicates regression to the 1st t Step's token, Indicates time step t All tokens before; This represents the softmax function. This represents the output layer weights of the base large language model; The expression for maximizing the likelihood probability of the target answer sequence based on the token prediction probability is: in, This represents the likelihood probability of the target answer sequence. This represents the total number of tokens for the target answer. Represents the natural logarithm. This represents all parameters of the projection layer and the base large language model. The first part represents the target answer. i One token, i This indicates the token index.

Images