Fish quantity evaluation method based on computer hearing and SENet network

Abstract

The application discloses a fish quantity evaluation method based on computer hearing and a SENet network, which can improve the accuracy of fish quantity evaluation and is especially suitable for a weak light environment. The evaluation method comprises the following steps: S10, collecting fish audio signals of different fish quantities to construct a first audio set; S20, performing frame division and windowing on the first audio set to obtain a second audio set; S30, converting the second audio set into spectrograms, screening the spectrograms based on a preset image quality standard to obtain an image set of spectrograms; S40, creating a fish quantity evaluation model based on a SENet network; an input layer parameter of the fish quantity evaluation model is a spectrogram in the image set of spectrograms, and an output layer parameter is a fish quantity; S50, training and testing the fish quantity evaluation model; and S60, estimating the quantity of a target fish group by using the fish quantity evaluation model.

Description

Technical Field

[0001] This invention relates to the field of fish farming technology, and in particular to a method for fish population assessment based on computer hearing and the SENet network. Background Technology

[0002] In fish farming, accurately assessing the number of fish in the pond is crucial for controlling appropriate stocking density, achieving precise feeding, and saving farming costs. It is a key parameter for achieving precise, efficient, and green fish farming.

[0003] In recent years, to improve aquaculture, computer vision technology has begun to be researched and initially applied in intelligent agriculture. However, there are many limitations to using computer vision technology for fish monitoring. For example, in turbid waters or under specific lighting conditions, such as when fish are densely clustered at the bottom of the pond or far from natural light sources, the images captured by the camera system will be distorted, affecting the accuracy of counting. Although infrared technology can work in dark environments, without appropriate infrared filtering devices, the clarity of the captured images may be insufficient, making it difficult to effectively assess the number of fish. Furthermore, mutual occlusion between individual fish also poses a challenge to the accurate estimation of fish numbers using computer vision technology.

[0004] Therefore, finding a more effective and intelligent method to assess fish populations and improve the accuracy of fish population prediction has become an urgent technical problem to be solved. Summary of the Invention

[0005] The purpose of this invention is to provide a fish population assessment method based on computer hearing and SENet network, which solves the problems existing in the prior art, improves the accuracy of fish population assessment, and is especially suitable for low light environments.

[0006] To achieve the above objectives, the solution of the present invention is:

[0007] A method for fish population assessment based on computer hearing and the SENet network includes the following steps:

[0008] Step S10: Collect audio signals of fish with different numbers of fish to construct the first audio set;

[0009] Step S20: Frame and window the first audio set to obtain the second audio set;

[0010] Step S30: Convert the second audio set into a spectrogram, and filter each spectrogram based on a preset image quality standard to obtain a spectrogram image set;

[0011] Step S40: Create a fish population assessment model based on the SENet network; the input layer parameters of the fish population assessment model are the spectrograms in the spectrogram image set, and the output layer parameters are the fish population.

[0012] Step S50: Train and test the fish population assessment model;

[0013] Step S60: Use the fish population assessment model to estimate the population of the target fish population.

[0014] Step S10 specifically involves:

[0015] A loudspeaker is placed at the bottom of the breeding tank against the wall, and several hydrophones are arranged opposite the loudspeaker. After different numbers of fish are released into the breeding tank in batches, white noise is played by the loudspeaker. The white noise is received by the hydrophones after passing through different numbers of fish, and the hydrophones collect the fish audio signals to construct the first audio set.

[0016] Step S20 specifically involves:

[0017] The first audio set is divided into frames with a unit of 1 second. During the audio framing process, the signal within each time window is windowed to obtain several audio segments, thereby constructing the second audio set.

[0018] Preferably, in step S20, framing includes a frame length, a frame shift, and a frame number;

[0019] The frame length is set to 1 second;

[0020] The formula for calculating the number of frames is as follows: ;in Indicates the number of frames. Indicates the total length of the audio signal. Indicates frame length. Indicates frame shift.

[0021] Preferably, in step S20, the window is a Hanning window, and the calculation formula for the Hanning window is as follows: ;in, Represents window functions, This indicates the position points of each sample in the window. Indicates the window length;

[0022] The process expression for windowing is as follows: ;in, This indicates the signal after windowing. This represents the original signal.

[0023] In step S40, the output layer, output range, and evaluation metrics of the SENet network are modified to adapt it to the image regression prediction task.

[0024] Modify the output layer into a fully connected layer with a single output node so that it outputs continuous predictions;

[0025] Add a linear activation function to the output range to ensure that the model output is any real value and is not restricted to a specific range.

[0026] In step S50, before training the fish population estimation model, the training data is normalized, and the coefficient of determination is... Mean absolute error Percentage error As a metric for evaluating the predictive performance of fish population assessment models on the test set;

[0027] Coefficient of determination The calculation formula is ;

[0028] Mean Absolute Error The calculation formula is ;

[0029] Percentage error The calculation formula is ;

[0030] above, Indicates the number of samples. Represents the true value. Indicates the predicted value. This represents the average of the sample size.

[0031] After adopting the above technical solution, the present invention has the following technical effects:

[0032] The fish population assessment model constructed by this invention is not easily affected by factors such as turbid water, uneven lighting, or mutual shading among individual fish in the aquaculture environment, thus improving the network's performance and generalization ability, exhibiting stronger robustness, and ultimately greatly enhancing the accuracy of fish population assessment. Attached Figure Description

[0033] Figure 1 This is a flowchart of a specific embodiment of the present invention;

[0034] Figure 2 This is a schematic diagram illustrating the loss changes during the training process of the fish population assessment model, which is a specific embodiment of the present invention.

[0035] Figure 3 This is a comparison chart of predicted and actual values ​​of the fish population assessment model in a specific embodiment of the present invention.

[0036] Figure 4 This is a graph showing the output fitting curve of the fish population assessment model in a specific embodiment of the present invention. Detailed Implementation

[0037] To further explain the technical solution of the present invention, the present invention will be described in detail below through specific embodiments.

[0038] The overall idea of ​​this invention is as follows: Audio data of different numbers of fish are collected using a hydrophone; several audio segments are obtained through frame segmentation and windowing; the audio data are converted into spectrograms, and each spectrogram is filtered based on image quality standards; the SENet network used for image classification and prediction tasks is modified to adapt to image regression prediction tasks; the training data used to train the fish population assessment model is normalized; a fish population assessment model is created using the SENet network adapted for image regression prediction tasks to learn the power intensity information in the spectrograms, making it less susceptible to the influence of factors such as turbid water, uneven lighting, or mutual occlusion among individual fish in the aquaculture environment, thus improving the network's performance and generalization ability, exhibiting stronger robustness, and ultimately greatly improving the accuracy of fish population assessment.

[0039] refer to Figure 1-4 As shown, this invention discloses a method for estimating fish populations based on computer hearing and the SENet network, comprising the following steps:

[0040] Step S10: Collect audio signals of fish with different numbers of fish to construct the first audio set;

[0041] Step S20: Frame and window the first audio set to obtain the second audio set;

[0042] Step S30: Convert the second audio set into spectrograms, and filter each spectrogram based on a preset image quality standard to obtain a spectrogram image set;

[0043] Step S40: Create a fish population assessment model based on the SENet network; the input layer parameters of the fish population assessment model are the spectrograms in the spectrogram image set, and the output layer parameters are the fish population.

[0044] Step S50: Train and test the fish population assessment model;

[0045] Step S60: Estimate the number of the target fish population using a fish population assessment model.

[0046] The specific extensions of the above steps are as follows:

[0047] The above step S10 is specifically as follows:

[0048] A loudspeaker is placed against the bottom wall of the rearing tank, and several hydrophones are positioned opposite it. White noise is played by the loudspeaker. After propagating through space, the white noise is received by the hydrophones. Because the white noise is attenuated to varying degrees by different numbers of fish, the sound signal received by the hydrophones shows a clear negative correlation with the number of fish. That is, the more fish there are, the weaker the sound intensity received by the hydrophones, and vice versa. Therefore, the number of fish can be monitored by the sound intensity received by the hydrophones. After different numbers of fish are introduced into the rearing tank in batches, white noise is played by the loudspeaker. The white noise is received by the hydrophones after passing through different numbers of fish, and the hydrophones collect the fish audio signals to construct the first audio set.

[0049] The above step S20 specifically involves: dividing the first audio set into frames with a unit of 1 second; during the audio framing process, windowing is applied to the signal within each time window to obtain several audio segments, thereby constructing the second audio set.

[0050] Furthermore, in step S20 above, framing includes a frame length, a frame shift, and a frame number;

[0051] The frame length is set to 1 second.

[0052] The above frame count calculation formula is as follows: ;in Indicates the number of frames. Indicates the total length of the audio signal. Indicates frame length. Indicates frame shift.

[0053] Meanwhile, in step S20 above, the window added is a Hanning window, and the calculation formula for the Hanning window is as follows: ;in, Represents window functions, This indicates the position points of each sample in the window. Indicates the window length;

[0054] The process expression for windowing is as follows: ;in, This indicates the signal after windowing. This represents the original signal.

[0055] In step S40 above, the output layer, output range, and evaluation metrics of the SENet network are modified to adapt it to the image regression prediction task:

[0056] Modify the output layer into a fully connected layer with a single output node so that it outputs continuous predictions;

[0057] Add a linear activation function to the output range to ensure that the model output is any real value and is not restricted to a specific range.

[0058] In step S50 above, before training the fish population estimation model, the training data is normalized, and the coefficient of determination is... Mean absolute error Percentage error As a metric for evaluating the predictive performance of fish population assessment models on the test set;

[0059] Coefficient of determination The calculation formula is ;

[0060] Mean Absolute Error The calculation formula is ;

[0061] Percentage error The calculation formula is ;

[0062] above, Indicates the number of samples. Represents the true value. Indicates the predicted value. This represents the average of the sample size.

[0063] The experimental data for this invention are as follows:

[0064] (1) Select eels for testing.

[0065] (2) During the training of the fish population assessment model, the model loss value was recorded after each epoch, and the trend of change was as follows: Figure 2 As shown. From Figure 2 It can be seen that the network loss decreases rapidly in the first 15 rounds, decreases slowly after 15 rounds, and basically stabilizes after 30 rounds, proving that the network model has converged.

[0066] (3) The trained fish population assessment model was tested, and the test results are as follows: Figure 3 , 4 As shown. From Figure 3 It can be seen that there are a total of 180 sets of test data. The difference between the predicted and actual values ​​is small for most samples, and the predicted and actual values ​​are basically consistent for a few samples. This shows that the fish population assessment model has good performance in regression prediction. Figure 4 The output fitting curve shown represents the fit between the predicted and actual values ​​obtained by the fish population assessment model during the testing phase. The correlation coefficient reaches 0.98, and the coefficient of determination is... The value is 0.96, indicating that the fish population assessment model constructed in this invention can accurately assess the population of eels in recirculating aquaculture systems. Its regression equation is... ;in Indicates the predicted value. Represents the actual value.

[0067] (4) Through the coefficient of determination Mean absolute error Percentage error The predictive performance of the fish population estimation model on the test set was evaluated based on the dimensional criteria, and the results are shown in the table below:

[0068]

[0069] As can be seen from the table above, the coefficient of determination of the fish population assessment model of the present invention is... The value reached 0.96, close to 1, indicating that it performed well in terms of fitting effect; the mean absolute error A value of 1.66 indicates that the prediction error has a relatively small range relative to the target variable. To more accurately evaluate the model's predictive performance, the mean absolute percentage error is used. As another key evaluation dimension, the result was 4.43%, which demonstrates the excellent performance of the fish population assessment model of the present invention in predicting the population and ensures the reliability of the prediction results.

[0070] The above embodiments and figures are not intended to limit the product form and style of the present invention. Any appropriate changes or modifications made by those skilled in the art should be considered as not departing from the patent scope of the present invention.

Claims

1. A method for estimating fish populations based on computer hearing and the SENet network, characterized in that... Includes the following steps: Step S10: Collect audio signals of fish with different numbers of fish to construct a first audio set; specifically, place a loudspeaker at the bottom wall of the breeding tank and arrange several hydrophones opposite the loudspeaker; after different numbers of fish are released into the breeding tank in batches, white noise is played by the loudspeaker, and the white noise is received by the hydrophones after passing through different numbers of fish, and the hydrophones collect the fish audio signals to construct the first audio set. Step S20: Frame and window the first audio set to obtain the second audio set; Step S30: Convert the second audio set into a spectrogram, and filter each spectrogram based on a preset image quality standard to obtain a spectrogram image set; Step S40: Create a fish population assessment model based on the SENet network; the input layer parameters of the fish population assessment model are the spectrograms in the spectrogram image set, and the output layer parameters are the fish population; wherein, by modifying the output layer, output range, and evaluation index of the SENet network, it is adapted to the image regression prediction task: the output layer is modified into a fully connected layer with a single output node so as to output continuous predicted values; a linear activation function is added to the output range to ensure that the model output is an arbitrary real value and is not restricted to a specific range; Step S50: Train and test the fish population assessment model; Step S60: Use the fish population assessment model to estimate the population of the target fish population.

2. The fish population assessment method based on computer hearing and SENet network as described in claim 1, characterized in that... Step S20 specifically involves: The first audio set is divided into frames with a unit of 1 second. During the audio framing process, the signal within each time window is windowed to obtain several audio segments, thereby constructing the second audio set.

3. The fish population assessment method based on computer hearing and SENet network as described in claim 2, characterized in that: In step S20, framing includes a frame length, a frame shift, and a frame number. The frame length is set to 1 second; The formula for calculating the number of frames is as follows: ;in Indicates the number of frames. Indicates the total length of the audio signal. Indicates frame length. Indicates frame shift.

4. The fish population assessment method based on computer hearing and SENet network as described in claim 2, characterized in that: In step S20, the window is a Hanning window, and the calculation formula for the Hanning window is as follows: ;in, Represents window functions, This indicates the position points of each sample in the window. Indicates the window length; The process expression for windowing is as follows: ;in, This indicates the signal after windowing. This represents the original signal.

5. The fish population assessment method based on computer hearing and SENet network as described in claim 1, characterized in that: In step S50, before training the fish population estimation model, the training data is normalized, and the coefficient of determination is... Mean absolute error Percentage error As a metric for evaluating the predictive performance of fish population assessment models on the test set; Coefficient of determination The calculation formula is ; Mean Absolute Error The calculation formula is ; Percentage error The calculation formula is ; above, Indicates the number of samples. Represents the true value. Indicates the predicted value. This represents the average of the sample size.

Images