Method and system for identifying and early warning dead broiler chickens based on image identification

Abstract

The invention provides a dead broiler recognition early warning method based on image recognition. The method comprises the steps that broiler breeding images are collected through a multispectral annular camera array; performing illumination compensation on the broiler breeding image to obtain an initial image; constructing a multi-branch feature extraction network; based on a multi-branch feature extraction network, feature extraction is carried out on the initial image through a multi-modal feature extraction algorithm, and broiler features are obtained; performing space-time joint reasoning according to the features of the broiler chickens to obtain a reasoning result; dynamically determining an early warning threshold value through the environment compensation factor; and executing different early warning processing operations according to a comparison result of the reasoning result and an early warning threshold value. According to the method, automatic identification of dead broiler chickens is realized through the constructed multi-branch feature extraction network and the dynamic threshold judgment mechanism, the identification accuracy is improved, the false alarm rate is reduced, and a reliable automatic inspection solution is provided for farms.

Description

Technical Field

[0001] The present invention relates to the technical field of image detection, and particularly to a method and system for identifying and warning dead broiler chickens based on image recognition. Background Art

[0002] With the development of modern poultry farming towards intensification and large-scale, the breeding density of broiler chickens continues to rise. In a closed environment with high temperature and humidity, the timely detection of individual deaths directly affects the efficiency of disease prevention and control and the economic benefits of farming. The traditional manual inspection method is limited by the operation frequency and visual fatigue, and it is difficult to achieve accurate positioning within 30 minutes after death, resulting in a sharp increase in the risk of cross-infection caused by the corruption of dead chickens.

[0003] The current mainstream technical solutions mainly use visible light cameras combined with infrared temperature measurement modules to detect static individuals through background difference method or optical flow method, supplemented by threshold temperature for judgment. However, there are still deficiencies, such as: lack of cross-modal feature alignment and fusion mechanism, resulting in the loss of weak abnormal responses; traditional network architectures are difficult to apply to special scenarios with sudden illumination changes and feather occlusion; the used threshold judgment mechanism is too rigid and the false alarm rate is too high. Therefore, it is very necessary to design a method and system for identifying and warning dead broiler chickens based on image recognition. Summary of the Invention

[0004] The purpose of the present invention is to provide a method and system for identifying and warning dead broiler chickens based on image recognition, which realizes the automatic identification of dead broiler chickens through the constructed multi-branch feature extraction network and dynamic threshold judgment mechanism, so as to improve the recognition accuracy and reduce the false alarm rate.

[0005] To achieve the above purpose, the present invention provides the following solutions:

[0006] A method for identifying and warning dead broiler chickens based on image recognition includes the following steps:

[0007] Collect broiler chicken breeding images through a multi-spectral annular camera array;

[0008] Perform light compensation on the broiler chicken breeding images to obtain initial images;

[0009] Construct a multi-branch feature extraction network; the multi-branch feature extraction network includes: a spatial feature branch, a temporal feature branch, a frequency domain feature branch, and a gating fusion module;

[0010] Based on the multi-branch feature extraction network, perform feature extraction on the initial images through a multi-modal feature extraction algorithm to obtain broiler chicken features;

[0011] Perform spatio-temporal joint reasoning based on the broiler chicken features to obtain a reasoning result;

[0012] Dynamically determine the early warning threshold through the environmental compensation factor;

[0013] Execute different early warning processing operations according to the comparison result between the inference result and the early warning threshold.

[0014] Optionally, the multi-spectral annular camera array is composed of an optical camera, a thermal imaging camera, and a depth sensing camera. The angle between any two of the optical camera, the thermal imaging camera, and the depth sensing camera is 120°, forming an annular arrangement.

[0015] Optionally, the expression for light compensation is: where I c (x,y) is the compensated pixel value, I(x,y) is the input pixel value, μ is the image mean, σ is the image standard deviation, t is the local time in hours, and sigmoid(·) is the sigmoid function.

[0016] Optionally, based on the multi-branch feature extraction network, perform feature extraction on the initial image through the multi-modal feature extraction algorithm to obtain broiler features, including:

[0017] Perform multi-scale feature extraction and adaptive receptive field selection on the initial image through the spatial feature branch to obtain optical image features;

[0018] Perform wavelet transform on the initial image through the temporal feature branch to obtain thermal imaging features;

[0019] Perform composite wavelet packet decomposition on the initial image through the frequency domain feature branch to obtain depth features;

[0020] Perform residual feature enhancement fusion operation on the optical image features, thermal imaging features, and depth features through the gated fusion module to obtain fusion features;

[0021] Perform feature distillation and spatio-temporal encoding on the fusion features to obtain broiler features.

[0022] Optionally, perform multi-scale feature extraction and adaptive receptive field selection on the initial image through the spatial feature branch to obtain optical image features, including:

[0023] Perform dynamic range compression on the initial image and use the bilateral filtering algorithm to eliminate the illumination mutation noise in the compressed image to obtain a denoised image;

[0024] Use the progressive dilation rate strategy to make the denoised image pass through five consecutive layers of dynamically dilated convolutional groups to obtain multi-scale features; the dynamically dilated convolutional groups are connected through adaptive activation units, and a second-order derivative constraint layer is set between the output end of the fourth layer of the dynamically dilated convolutional group and the adaptive activation unit;

[0025] After performing global average pooling and max pooling on the multi-scale features respectively, two outputs are obtained. After concatenating the two outputs through the SENet architecture, they pass through two fully connected layers in sequence to obtain the optical image features.

[0026] Optionally, the initial image is decomposed by complex wavelet packets through the frequency domain feature branch to obtain depth features, including:

[0027] The distances between each pixel point in the initial image and the fitting plane are obtained through the RANSAC algorithm, and screening and Gaussian filtering operations are performed to obtain optimized features;

[0028] The optimized features are decomposed by wavelet packets for 5 layers to obtain decomposed features;

[0029] The decomposed features are reconstructed to the spatial domain through inverse wavelet packet transform to obtain enhanced features;

[0030] The enhanced depth features are convolved through a 3×3×3 convolutional kernel to obtain depth features.

[0031] Optionally, feature distillation and spatio-temporal encoding are performed on the fusion features to obtain broiler features, including:

[0032] The fusion features are double-filtered according to the spatial saliency map obtained by using the Grad-CAM method and the channel importance coefficients obtained by using the global covariance matrix to obtain distilled features;

[0033] The fusion features and the distilled features are concatenated through a bidirectional LSTM network to obtain multi-granularity features;

[0034] The distilled features and the multi-granularity features are fused through a cross-attention mechanism to obtain broiler features.

[0035] Optionally, spatio-temporal joint reasoning is performed according to the broiler features to obtain reasoning results, including:

[0036] Death determination indicators are obtained according to the broiler features; the death determination indicators include: posture anomaly index, thermodynamic instability coefficient, and motion chaos index;

[0037] The theoretical probability of death is calculated through the death determination indicators; the calculation formula of the theoretical probability of death is:

[0038] P(m) is the theoretical probability of death, m is the continuous monitoring time, η(m) is the time parameter, k(m) is the risk shape parameter, β 1 、β 2 and β 3 are the posture anomaly weight, the thermal instability weight, and the motion chaos weight respectively, β 1 +β 2+β 3 = 1, where PA, TS, and MC are the posture anomaly index, the thermodynamic instability coefficient, and the motion chaos index respectively, and exp(·) is the natural exponential function;

[0039] The loss optimization of the theoretical probability of death is performed through the loss function to obtain the predicted probability of death, and the predicted probability of death is regarded as the inference result.

[0040] Optionally, the calculation formula for the warning threshold is: where Y is the warning threshold, θ t is the theoretical threshold, RH is the environmental humidity, and T env is the chicken house temperature.

[0041] A dead broiler recognition and warning system based on image recognition, comprising:

[0042] An information collection module for collecting broiler breeding images through a multi-spectral annular camera array;

[0043] An image processing module for performing light compensation on the broiler breeding images to obtain initial images;

[0044] A network construction module for constructing a multi-branch feature extraction network;

[0045] A feature extraction module for extracting broiler features from the initial images through a multi-modal feature extraction algorithm based on the multi-branch feature extraction network;

[0046] A feature inference module for performing spatio-temporal joint inference based on the broiler features to obtain an inference result;

[0047] A dynamic compensation module for dynamically determining the warning threshold through an environmental compensation factor;

[0048] A recognition and warning module for performing different warning processing operations according to the comparison result between the inference result and the warning threshold.

[0049] According to the specific embodiments provided by the present invention, the following technical effects are disclosed: The method for identifying and warning dead broiler chickens based on image recognition provided by the present invention includes: collecting broiler chicken breeding images through a multi-spectral annular camera array; performing light compensation on the broiler chicken breeding images to obtain initial images; constructing a multi-branch feature extraction network; based on the multi-branch feature extraction network, performing feature extraction on the initial images through a multi-modal feature extraction algorithm to obtain broiler chicken features; performing spatio-temporal joint reasoning based on the broiler chicken features to obtain a reasoning result; dynamically determining a warning threshold through an environmental compensation factor; and performing different warning processing operations according to the comparison result between the reasoning result and the warning threshold. This method realizes the automatic identification of dead broiler chickens through the constructed multi-branch feature extraction network and dynamic threshold judgment mechanism, improves the accuracy of identification and reduces the false alarm rate, providing a reliable automatic inspection solution for the breeding farm. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention, and those of ordinary skill in the art can also obtain other drawings based on these drawings without creative efforts.

[0051] Figure 1 It is a flowchart for identifying and warning dead broiler chickens of the present invention;

[0052] Figure 2 It is a flowchart for feature extraction of the present invention;

[0053] Figure 3 It is a flowchart for spatio-temporal joint reasoning of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0054] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all of them. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.

[0055] To make the above objects, features, and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below with reference to the drawings and specific embodiments.

[0056] As Figure 1 shown, the present invention provides a method for identifying and warning dead broiler chickens based on image recognition, including the following steps:

[0057] Step 100: Collect images of broiler chicken farming using a multi-spectral annular camera array;

[0058] Specifically, the multi-spectral annular camera array consists of a visible light camera, a thermal imaging camera, and a depth sensing camera. The angle between any two of the visible light camera, the thermal imaging camera, and the depth sensing camera is 120°, forming an annular arrangement. The camera acquisition period is 30 seconds per time.

[0059] Step 200: Perform light compensation on the images of broiler chicken farming to obtain initial images;

[0060] Specifically, the expression for light compensation is:

[0061]

[0062] where, I c (x,y) is the compensated pixel value, I(x,y) is the input pixel value, μ is the image mean, σ is the image standard deviation, t is the local time in hours, and sigmoid(·) is the sigmoid function.

[0063] Step 300: Construct a multi-branch feature extraction network; the multi-branch feature extraction network includes: a spatial feature branch, a temporal feature branch, a frequency domain feature branch, and a gating fusion module;

[0064] Specifically, the spatial feature branch consists of 5 layers of dilated convolutional groups. The first layer consists of a 3×3 standard convolution with 64 channels and a dilation factor of 1 and an adaptive activation unit. The second layer consists of a dilated convolution with 128 channels and a dilation factor of 2 and an adaptive activation unit. The third layer consists of a dilated convolution with 256 channels and a dilation factor of 4 and an adaptive activation unit. The fourth layer consists of a dilated convolution with 512 channels and a dilation factor of 8, a second derivative constraint layer, and an adaptive activation unit. The fifth layer consists of a dilated convolution with 1024 channels and a dilation factor of 16 and a weighting unit. The second derivative constraint algorithm of the second derivative constraint layer is the edge enhancement algorithm of the Hessian matrix. The spatial feature branch adopts a dynamic dilation mechanism, enabling the dilation rate of each layer to increase exponentially, capable of capturing cross-scale features from local textures to global structures, and improving the pose estimation accuracy in dense scenes while ensuring high resolution.

[0065] The temporal feature branch consists of a temporal encoding layer and a thermodynamic differential module. The temporal encoding layer consists of a bidirectional LSTM network with 256 hidden units and a time window of 8 and a temporal convolutional network with 4 layers of dilated convolution. The thermodynamic differential module is used to calculate the gradient of the temperature field. The detection sensitivity to the weak movements of dying broiler chickens is improved through the dependence of the bidirectional LSTM network on time.

[0066] The frequency-domain feature branch consists of a decomposition layer and a frequency-band optimization unit. The decomposition layer uses the db8 wavelet basis, with 5 decomposition levels, and performs wavelet decomposition through the optimal basis selection criterion. The frequency-band optimization unit preferentially retains the high-information sub-bands with concentrated energy and kurtosis > 3 in the decomposed data through the energy weighting and adaptive threshold mechanism.

[0067] Step 400: Based on the multi-branch feature extraction network, use the multi-modal feature extraction algorithm to extract features from the initial image to obtain broiler features; the specific steps are as Figure 2 shown, including:

[0068] Step 401: Perform multi-scale feature extraction and adaptive receptive field selection on the initial image through the spatial feature branch to obtain optical image features; including:

[0069] Compress the dynamic range of the initial image and use the bilateral filtering algorithm to eliminate the illumination mutation noise in the compressed image to obtain a denoised image;

[0070] Use the progressive dilation rate strategy to make the denoised image pass through five consecutive layers of dynamically expanding convolutional groups to obtain multi-scale features; after each convolution, an adaptive activation unit is connected, and the calculation formula for its activation threshold α is: α = 0.2 + 0.1×log(1 + local variance); the expression of the progressive dilation rate strategy is:

[0071]

[0072] where k is the convolution kernel radius, l = 0, 1, 2, 3, 4 is the dilation rate level, φ(z) = z·tanh(z), tanh(·) is the activation function, V t is the denoised image, W l (i,j) is the weight parameter of the l-th layer convolution kernel, (i,j) is the convolution kernel position coordinate, and (x,y) is the pixel position coordinate.

[0073] Perform global average pooling and max pooling on the multi-scale features respectively to obtain two outputs, splice the two outputs through the SENet architecture, and then pass through two fully connected layers in sequence to obtain optical image features. Among them, a spatial attention mechanism is used in the splicing process, and its spatial weight is obtained by generating an offset field through deformable convolution, which is used to dynamically adjust the receptive field shape of each position on the feature map. Capture multi-scale information from local details to global semantics through progressive dilation convolution, so that the obtained high-dimensional optical image features provide basic spatial context information for subsequent fusion.

[0074] Step 402: Perform wavelet transform on the initial image through the time feature branch to obtain thermal imaging features;

[0075] Specifically, first adopt a multi-resolution registration strategy. Calculate the translation matrix between adjacent frames at 1 / 4 resolution, and then perform an affine transformation at the original resolution. After obtaining the compensated thermal imaging sequence, perform differential processing on the temperature change rate field obtained by calculating five consecutive frames of images. Then perform a two-dimensional discrete wavelet transform on the temperature change rate field, and select the db6 wavelet basis for three-layer decomposition. In each layer of decomposition, perform denoising on the approximation coefficients based on the local entropy threshold. The calculation formula for the local entropy threshold is: τ = μ + 0.5σ, where μ is the sub-band mean and σ is the standard deviation. Finally, perform dynamic feature encoding through a hybrid architecture composed of a TCN network and an LSTM network, using the local temperature gradient of the current frame as the input to output the thermal imaging features. The expression of the overall process is as follows:

[0076]

[0077]

[0078] i t = σ(W i *[T t ,c t-1 );

[0079] o t = σ(W o *[FFT(T t ),h t-1 );

[0080]

[0081]

[0082] Among them, T t is the thermal imaging image of the t-th frame, * is the convolution operation, W c , W f , W i and W o are the weights of different convolution kernels, is the spatial gradient of the temperature field, is the time modulation factor, Re(·) is the operation of taking the real part, is the wavelet transform operation, ⊙ is the element-wise multiplication, σ(·) is the sigmoid function, LayerNorm(·) is the layer normalization operation, FFT(·) is the fast Fourier transform.

[0083] Step 403: Perform composite wavelet packet decomposition on the initial image through the frequency domain feature branch to obtain deep features;

[0084] Specifically, first within a 5×5 sliding window, the distances between each pixel point in the initial image and the fitted plane are obtained through the RANSAC algorithm and regarded as confidence levels. Pixel points with confidence levels lower than 0.7 are removed, and then Gaussian filtering is used for restoration to obtain optimized features. Then, the optimized features are decomposed by the optimal basis selection algorithm using 5-layer wavelet packet decomposition. During each layer of decomposition, the weighted sum of the entropy value and energy value of each node is calculated, and the decomposition path with the minimum weighted sum is selected to obtain the optimized subband coefficients. Next, the optimized subband coefficients are reconstructed to the spatial domain through inverse wavelet packet transform to obtain enhanced features. Finally, the enhanced depth features are convolved using a 3×3×3 convolutional kernel to obtain depth features.

[0085] It should be noted that the composite wavelet packet decomposition used in the frequency domain feature branch improves the accuracy of extracting texture and structural features in depth information, making up for the perceptual limitations of visible light and thermal imaging.

[0086] Step 404: Perform a residual feature enhancement fusion operation on the optical image features, thermal imaging features, and depth features through a gated fusion module to obtain fusion features;

[0087] Specifically, first a spatial deformation field is generated through deformable convolution to unify the spatial resolutions of the thermal imaging features and depth features to the scale of the optical image features. Then, a channel-space-time three-dimensional attention gate is constructed to perform weighted fusion on the optical image features, thermal imaging features, and depth features. Finally, the features after weighted fusion are connected with the optical image features through cross-layer residual connection. Before the residual connection, the optical image features are adaptively normalized, and a random path dropout strategy is adopted during the residual connection process to prevent overfitting.

[0088] More specifically, the channel allocation coefficient, spatial weight, and temporal attention of the channel-space-time three-dimensional attention gate are obtained through softmax normalization operation, parallel dilated convolution branches, and autocorrelation matrix distribution respectively.

[0089] It should be noted that the gated fusion module realizes the precise alignment and adaptive weighting of cross-modal features through a three-dimensional attention mechanism, and the residual connection ensures that key information is not lost, guaranteeing the integrity of information.

[0090] Step 405: Perform feature distillation and spatio-temporal encoding on the fusion features to obtain broiler features. This includes:

[0091] The fusion features are double-filtered according to the spatial saliency map obtained using the Grad-CAM method and the channel importance coefficient obtained using the global covariance matrix to obtain distilled features;

[0092] The fusion features and distilled features are concatenated through a bidirectional LSTM network to obtain multi-granularity features;

[0093] The distilled features and multi-granularity features are fused through a cross-attention mechanism to obtain broiler features. Among them, the distilled features are used as query vectors, and the multi-granularity features are used as key-value pairs.

[0094] Step 500: Perform spatio-temporal joint reasoning based on the broiler features to obtain an inference result; the specific steps are as Figure 3 shown, including:

[0095] Step 501: Obtain a death determination index based on the broiler features; the death determination index includes: posture anomaly index, thermodynamic instability coefficient, and motion chaos index;

[0096] Specifically, the calculation formula for the posture anomaly index is:

[0097]

[0098] Among them, is the weight of the i-th joint, θ i is the radian of the i-th joint, is the angle reference value of a healthy broiler, λ p is the time decay factor, which is 0.05 in this embodiment, Δt i is the duration of the abnormal state. The 5 joints are the neck, both wings, and both legs.

[0099] The calculation formula for the thermodynamic instability coefficient is:

[0100]

[0101] Among them, T is the thermal imaging pixel temperature, t is the current time, τ is the time integration window, which is 300s in this embodiment, T g is the temperature field gradient norm, max(·) is the maximum value operation, and ξ is a time point within the integration interval, representing the time dimension variable for differentiating the temperature field during the integration process.

[0102] The calculation formula for the motion chaos index is:

[0103]

[0104] Among them, N is the number of key body parts, which is 5 in this embodiment, v i is the actual optical flow velocity vector of the i-th part (obtained through the dense optical flow algorithm), is the predicted velocity vector of the i-th part (obtained by temporal prediction using the LSTM model), σ v is the velocity normalization factor, Δφ k is the motion direction deviation angle, φ t is the angle deviation threshold, which is π / 4 in this embodiment, It is an indicator function, which takes 1 when the condition in the brackets is met and 0 otherwise.

[0105] Step 502: Calculate the theoretical probability of death through the death determination index;

[0106] Specifically, the calculation formula of the theoretical probability of death includes:

[0107]

[0108] η(m) = η 0 ·exp(-α η ·[PA + TS + MC]);

[0109]

[0110] where P(m) is the theoretical probability of death, m is the continuous monitoring time, η(m) is the time parameter, k(m) is the risk shape parameter, β 1 , β 2 and β 3 are the attitude anomaly weight, the thermal instability weight, and the motion chaos weight respectively, β 1 +β 2 +β 3 = 1, PA, TS, and MC are the attitude anomaly index, the thermodynamic instability coefficient, and the motion chaos index respectively, exp(·) is the natural exponential function, η 0 is the reference time parameter, which is 600s in this embodiment, k 0 is the initial value of the shape parameter, which is 1.2 in this embodiment, α η is the coupling coefficient of PA and MC, which is 0.15 in this embodiment, and γ is the frequency domain memory decay coefficient, which is [0.1, 0.08, 0.05] in this embodiment.

[0111] Step 503: Optimize the loss of the theoretical probability of death through the loss function to obtain the predicted probability of death, and regard the predicted probability of death as the inference result.

[0112] Specifically, the loss function is obtained by performing a weighted average operation on binary cross-entropy, Kullback-Leibler divergence, and regularization loss.

[0113] Step 600: Dynamically determine the warning threshold through the environmental compensation factor;

[0114] Specifically, the calculation formula of the warning threshold is:

[0115]

[0116] where Y is the warning threshold, θ tis the theoretical threshold, RH is the environmental humidity, and T env is the temperature of the chicken coop.

[0117] Step 700: Perform different early warning processing operations according to the comparison result between the inference result and the early warning threshold.

[0118] Specifically, when P(m)>Y is satisfied within 3 consecutive detection cycles (90 seconds), a yellow early warning is triggered; when the yellow early warning persists for 2 cycles without being lifted, an orange early warning is triggered; when P(m)>1.2Y is satisfied within 5 cumulative cycles or the orange early warning persists for 2 cycles without being lifted, a red early warning is triggered; when P(m)<0.8Y for 3 consecutive cycles or after manual reset, the early warning is lifted.

[0119] More specifically, when a yellow early warning is triggered, start the negative pressure ventilation system in the area where the recognition target is located for ventilation; when an orange early warning is triggered, start the rotating stroboscopic warning light and wake up the robotic arm, and at the same time, turn off the automatic feeder in the area where the recognition target is located; when a red early warning is triggered, the robotic arm locates and grabs the recognition target based on computer vision.

[0120] The present invention also provides a dead broiler recognition and early warning system based on image recognition, including:

[0121] An information collection module for collecting broiler breeding images through a multi-spectral annular camera array;

[0122] An image processing module for performing light compensation on the broiler breeding images to obtain initial images;

[0123] A network construction module for constructing a multi-branch feature extraction network;

[0124] A feature extraction module for extracting broiler features from the initial images through a multi-modal feature extraction algorithm based on the multi-branch feature extraction network;

[0125] A feature inference module for performing spatio-temporal joint inference based on the broiler features to obtain an inference result;

[0126] A dynamic compensation module for dynamically determining the early warning threshold through an environmental compensation factor;

[0127] An identification and early warning module for performing different early warning processing operations according to the comparison result between the inference result and the early warning threshold.

[0128] The beneficial effects of the present invention are as follows:

[0129] 1) Through multi-spectral image fusion and spatio-temporal joint inference, accurate identification of dead broilers is achieved, and the detection accuracy is greatly improved;

[0130] 2) Adaptive fusion of visible light, thermal imaging, and depth information is achieved through an innovative multi-branch feature extraction network, solving the interference problems of feather occlusion and sudden illumination changes and reducing the recognition error rate;

[0131] 3) By introducing an environmental compensation factor into the dynamic threshold adjustment mechanism, the detection stability in extreme environments is improved, solving the limitations of traditional fixed-threshold methods;

[0132] 4) Through a real-time dynamic warning method, precise hierarchical disposal from local ventilation adjustment to full-field emergency response is achieved, while reducing equipment energy consumption and false action rates.

[0133] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. For the same and similar parts among the various embodiments, reference can be made to each other.

[0134] Specific examples are used in the present invention to elaborate on the principles and implementation manners of the present invention. The descriptions of the above embodiments are only used to help understand the method and its core idea of the present invention; at the same time, for those of ordinary skill in the art, based on the idea of the present invention, there will be changes in the specific implementation manners and application scopes. In summary, the content of this specification should not be construed as a limitation to the present invention.

Claims

1. A dead broiler chicken identification and early warning method based on image recognition, characterized in that: The steps include: Collect broiler farming images through a multispectral annular camera array; Performing illumination compensation on the broiler breeding image to obtain an initial image; Constructing a multi-branch feature extraction network; the multi-branch feature extraction network includes: a spatial feature branch, a temporal feature branch, a frequency domain feature branch and a gated fusion module; Based on the multi-branch feature extraction network, extracting features from the initial image by a multimodal feature extraction algorithm to obtain broiler features; Performing spatiotemporal joint reasoning according to the broiler chicken characteristics to obtain a reasoning result; Dynamically determine the warning threshold through environmental compensation factors; Different warning processing operations are performed according to the comparison result between the inference result and the warning threshold.

2. The dead broiler chicken identification and early warning method based on image recognition according to claim 1 is characterized in that: The multi-spectral annular camera array is composed of an optical camera, a thermal imaging camera and a depth sensing camera. The angle between any two of the optical camera, the thermal imaging camera and the depth sensing camera is 120°, forming a circular arrangement.

3. The dead broiler chicken identification and early warning method based on image recognition according to claim 1 is characterized in that: The expression of the illumination compensation is: Among them, I c (x, y) is the compensated pixel value, I(x, y) is the input pixel value, μ is the image mean, σ is the image standard deviation, t is the local time hour, and sigmoid(·) is the sigmoid function.

4. The dead broiler chicken identification and early warning method based on image recognition according to claim 1 is characterized in that: Based on the multi-branch feature extraction network, the initial image is subjected to feature extraction by a multimodal feature extraction algorithm to obtain broiler features, including: Performing multi-scale feature extraction and adaptive receptive field selection on the initial image through the spatial feature branch to obtain optical image features; Performing wavelet transform on the initial image through a time feature branch to obtain thermal imaging features; Performing composite wavelet packet decomposition on the initial image through a frequency domain feature branch to obtain a depth feature; Performing a residual feature enhancement fusion operation on the optical image feature, the thermal imaging feature and the depth feature through a gated fusion module to obtain a fusion feature; The fused features are subjected to feature distillation and spatiotemporal encoding to obtain the broiler features.

5. The dead broiler chicken identification and early warning method based on image recognition according to claim 4 is characterized in that: The spatial feature branch is used to perform multi-scale feature extraction and adaptive receptive field selection on the initial image to obtain optical image features, including: The initial image is subjected to dynamic range compression, and a bilateral filtering algorithm is used to eliminate sudden illumination noise of the compressed image to obtain a de-noised image; A progressive dilation rate strategy is used to obtain multi-scale features of the denoised image through five layers of dynamically dilated convolution groups connected in sequence; the dynamically dilated convolution groups are connected through adaptive activation units, and a second-order derivative constraint layer is provided between the output end of the fourth layer of the dynamically dilated convolution group and the adaptive activation unit; The multi-scale features are subjected to global average pooling and maximum pooling respectively to obtain two outputs, and the two outputs are spliced ​​through the SENet architecture and then sequentially passed through two fully connected layers to obtain the optical image features.

6. The dead broiler identification and early warning method based on image recognition according to claim 4 is characterized in that: The initial image is subjected to composite wavelet packet decomposition through the frequency domain feature branch to obtain a depth feature, including: The distance between each pixel in the initial image and the fitting plane is obtained by using the RANSAC algorithm, and screening and Gaussian filtering operations are performed to obtain optimized features; Performing 5-layer wavelet packet decomposition on the optimized features to obtain decomposed features; Reconstructing the decomposed features into the spatial domain by inverse wavelet packet transform to obtain enhanced features; The enhanced depth feature is convolved with a 3×3×3 convolution kernel to obtain the depth feature.

7. The dead broiler chicken identification and early warning method based on image recognition according to claim 4 is characterized in that: The fusion features are subjected to feature distillation and spatiotemporal encoding to obtain the broiler features, including: Double filtering the fused features according to the spatial saliency map obtained by using the Grad-CAM method and the channel importance coefficient obtained by using the global covariance matrix to obtain distilled features; The fusion feature and the distillation feature are concatenated through a bidirectional LSTM network to obtain a multi-granularity feature; The distilled features and the multi-granularity features are fused through a cross-attention mechanism to obtain the broiler features.

8. The dead broiler chicken identification and early warning method based on image recognition according to claim 1 is characterized in that Based on the broiler chicken characteristics, spatiotemporal joint reasoning is performed to obtain reasoning results, including: Obtaining death determination indicators according to the broiler characteristics; the death determination indicators include: posture abnormality index, thermodynamic instability coefficient and motion chaos index; The theoretical probability of death is calculated by the death determination index; the calculation formula of the theoretical probability of death is: Where P(m) is the theoretical probability of death, m is the continuous monitoring time, η(m) is the time parameter, k(m) is the risk shape parameter, β1, β2 and β3 are the posture abnormality weight, thermal instability weight and motion chaos weight, β1+β2+β3=1, PA, TS and MC are the posture abnormality index, thermodynamic instability coefficient and motion chaos index, respectively, and exp(·) is the natural exponential function; The theoretical probability of death is optimized through a loss function to obtain a predicted probability of death, and the predicted probability of death is regarded as the inference result.

9. The dead broiler chicken identification and early warning method based on image recognition according to claim 1 is characterized in that: The calculation formula of the warning threshold is: Among them, Y is the warning threshold, θ t is the theoretical threshold, RH is the ambient humidity, T env The temperature of the chicken house.

10. A dead broiler identification and early warning system based on image recognition, characterized in that: include: An information collection module, used to collect broiler farming images through a multi-spectral annular camera array; An image processing module, used for performing illumination compensation on the broiler breeding image to obtain an initial image; Network construction module, used to build a multi-branch feature extraction network; A feature extraction module, used to extract features from the initial image through a multimodal feature extraction algorithm based on the multi-branch feature extraction network to obtain broiler features; A feature reasoning module, used for performing spatiotemporal joint reasoning according to the broiler features to obtain a reasoning result; A dynamic compensation module is used to dynamically determine the warning threshold through environmental compensation factors; The identification and warning module is used to perform different warning processing operations according to the comparison result between the inference result and the warning threshold.

Images