Unmanned boat feeding method for target fish schools based on temporal network algorithm

Abstract

The present invention relates to the field of feeding technology in underwater fish management, and in particular to an unmanned boat feeding method for a target fish school based on a temporal network algorithm, so as to solve the problems of fish school identification, position distance positioning and classification and feeding of different fish schools in an underwater environment. By treating each fish in the fish school as a temporal node, extracting characteristic data of swimming speed, pattern spot category, spacing and pixel ratio of fish on an image, constructing a characteristic sequence of continuous fish, and providing strong data support for temporal network analysis, the improved LSTM network analysis can accurately identify the type of fish school and reasonably select suitable feed. At the same time, according to the movement speed of the fish school, the number of input vectors is intelligently adjusted to achieve dynamic adaptation. With the help of an adaptive amplification factor, the network algorithm will automatically adjust according to the classification accuracy of the pattern spots, and further improve the system accuracy. In addition, through a sonar device, the distance between the ship and the fish is accurately measured to accurately guide the movement of the ship and implement the feeding operation.

Description

Technical Field

[0001] The invention relates to the technical field of feeding in underwater fish management, and in particular to an unmanned boat feeding method for a target fish school based on a temporal network algorithm. Background Art

[0002] At present, in the field of underwater fish management, fish identification, positioning and classified feeding are crucial to underwater fish management. However, existing technologies cannot solve the problems of how to identify, locate and classified feed fish, which brings significant challenges to underwater fish management.

[0003] Existing technologies often lack accurate fish feature extraction, resulting in insufficient recognition accuracy. At the same time, since the timing and movement speed of fish schools are usually not taken into account, the feeding is not accurate enough. In addition, the distance measurement technology is also unstable, making it difficult to achieve accurate feeding over long distances. Summary of the invention

[0004] The purpose of the present invention is to provide an unmanned boat feeding method for target fish schools based on a temporal network algorithm. Based on unmanned boat equipment and innovatively introducing a temporal network algorithm, the fish schools are regarded as time series nodes, and the swimming speed, pattern spot category, pixel ratio and other characteristics are combined to overcome the previous recognition defects; the adaptive adjustment of the input quantity and the amplification factor adjustment improves the system adaptability and accuracy; and the distance measurement based on sonar technology ensures the accuracy of feeding, which provides a theoretical basis for realizing intelligent underwater feeding, so as to solve the technical problems existing in the above-mentioned background technology.

[0005] To achieve the above object, the technical solution of the present invention is as follows:

[0006] A method for feeding a target fish school by an unmanned boat based on a temporal network algorithm specifically comprises the following steps:

[0007] Step 1: Preliminary collection of target fish data;

[0008] Step 2: Extracting target fish data features;

[0009] Step 3: Construction of temporal network;

[0010] Step 4: Ship movement and feed feeding.

[0011] Furthermore, the preliminary collection of target fish school data is specifically as follows: data is collected for each fish in the target fish school, wherein the data collection object is ten consecutive fish in the target fish school, and the swimming speed of these ten consecutive fish and the position of the target fish school are specifically collected.

[0012] Furthermore, the swimming speeds of ten consecutive fish in the target fish school are collected specifically as follows: the swimming speeds of ten consecutive fish in the target fish school are obtained by using radar to measure the frequency and phase changes of the echo signal to obtain the swimming speed of each fish in the target fish school.

[0013] Furthermore, the collection of the position data of the target fish school is specifically: using an underwater high-definition camera to collect images of the fish school position scanned by the radar to obtain a picture of the target fish school.

[0014] Furthermore, the target fish data feature extraction is specifically as follows: feature extraction is performed on the target fish picture, and the extracted features include: fish pattern spot category and size spacing ratio between fish, wherein the fish pattern spot category is extracted using a network architecture based on YOLO transfer learning, the fish category is set to 1, and the remaining options are all 0; in addition, the regression frame detection part of the YOLO algorithm is intercepted, and a prediction frame is output, and the fish spacing is extracted through the prediction frame and according to the size spacing ratio formula.

[0015] Furthermore, the size spacing ratio formula is:

[0016]

[0017] Among them, Sp is the size spacing ratio between two adjacent fish, Ce1 and Ce2 are the center points of the target detection frames of the two adjacent fish, X2 and X1 are the upper right corner horizontal coordinate value and the upper left corner coordinate value of the target detection frame of the right fish, and X′2 and X1′ are the upper right corner horizontal coordinate value and the upper left corner coordinate value of the target detection frame of the left fish.

[0018] Furthermore, the construction of the temporal network is specifically as follows: the feature vector of each fish in the target fish school is regarded as an input vector in the temporal network, wherein the feature vectors are: fish pattern spot category, size spacing ratio between fish and swimming speed of each fish; the Del-Act activation function is used to determine the number of input vectors, the Random decision operator is used to determine the starting point of the input vector, and finally a static fish recognition adaptive amplification factor Factor is proposed to determine the number of extractions of the Random operator.

[0019] Furthermore, the Del-Act activation function is:

[0020]

[0021] Where num represents the number of feature vectors to be input, which is a maximum of 10; v i Represented as the speed value of the i-th fish;

[0022] The Random decision operator is:

[0023] Random(1,8-num);

[0024] Among them, num is the number of feature vectors to be input, and Random(1,8-num) means randomly selecting a positive integer from 1 to 8-num, including 1 and 8-num;

[0025] The adaptive amplification factor Factor is:

[0026]

[0027] Among them, Factor is the adaptive amplification factor, Arc(class) is the accuracy of fish pattern spot classification, and the entire network can only operate when the value of Arc(class) is greater than 0.9.

[0028] Furthermore, the ship movement and feed feeding are specifically as follows: through the sonar device on the unmanned ship, sound waves are emitted underwater to measure the time it takes for the sound waves to reach the fish and return from the ship, and the distance is calculated. Then the unmanned ship sails to the location of the fish school and feeds the fish accurately.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] 1. The present invention adopts a temporal network algorithm, comprehensively considers the characteristics and movement patterns of fish, and realizes accurate identification, positioning and classified feeding of fish schools; the system can accurately distinguish different fish and provide suitable feed, thereby optimizing the management effect.

[0031] 2. The present invention can dynamically adapt to the movement speed and behavior patterns of different fish schools by adaptively adjusting the number of input vectors and the amplification factor; this makes the feeding process more flexible and accurate, and improves the adaptability of the system.

[0032] 3. The present invention introduces sonar technology to measure the distance between the ship and the fish to ensure the accuracy of feeding; the system can accurately guide the movement of the ship based on real-time distance information, so as to implement feeding at the correct position, avoiding the errors in traditional methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 is a flow chart of the present invention;

[0034] Figure 2 is a schematic diagram of a school of fish image in the present invention;

[0035] Figure 3 It is a schematic diagram of the fish target detection network architecture in the present invention;

[0036] Figure 4 It is a schematic diagram of the sequential network architecture in the present invention. DETAILED DESCRIPTION

[0037] The present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments.

[0038] See also Figure 1 As shown, the unmanned boat feeding method of the target fish school based on the temporal network algorithm specifically includes the following steps:

[0039] Step 1: Preliminary data collection of target fish school; In this step, data collection and characterization work needs to be performed on each fish in the target fish school. Among them, the object of data collection is ten consecutive fish in the target fish school, and the swimming speed of these ten consecutive fish and the position of the target fish school are specifically collected. The movement speed of the ten consecutive fish in the target fish school is measured by radar through measuring the frequency and phase changes of the echo signal to obtain the swimming speed of each fish in the target fish school. The underwater high-definition camera is used to collect images of the fish school position scanned by the radar to obtain the target fish school picture, such as Figure 2 shown.

[0040] Step 2: Extracting features of target fish data; the preliminary data of the target fish obtained in step 1 include: the swimming speed of each fish in the target fish and the target fish picture. In this step, feature extraction is performed on the target fish picture, and the extracted features include: fish pattern spot category and the size spacing ratio between fish. The fish pattern spot category is extracted using a network architecture based on YOLO transfer learning, and the fish category is set to 1, and the remaining options are all 0; in addition, the regression frame detection part of the YOLO algorithm is intercepted, and the prediction frame is output, and the fish spacing is extracted through the prediction frame and according to the size spacing ratio formula.

[0041] The formula for the size spacing ratio in step 2 is:

[0042]

[0043] Among them, Sp is the size spacing ratio between two adjacent fish, Ce1 and Ce2 are the center points of the target detection frames of the two adjacent fish, X2 and X1 are the upper right corner horizontal coordinate value and the upper left corner coordinate value of the target detection frame of the right fish, and X′2 and X1′ are the upper right corner horizontal coordinate value and the upper left corner coordinate value of the target detection frame of the left fish.

[0044] The size spacing ratio between adjacent fish can be calculated using the size spacing ratio formula, such as Figure 3As shown in the figure, Ce1 and Ce2 represent the center point values ​​of the target detection frames of Fish-1 and Fish-2 respectively, X2 and X1 are the upper right corner horizontal coordinate value and the upper left corner coordinate value of the target detection frame of Fish-2 respectively, X′2 and X1′ are the upper right corner horizontal coordinate value and the upper left corner coordinate value of the target detection frame of Fish-1 respectively, and the size spacing ratio between Fish-1 and Fish-2 can be calculated by the size spacing ratio formula.

[0045] In addition, a strong convolution operation is performed on the individual contents in the detection frame obtained by migrating the YOLO network framework, where the Conv operator is 3*3, the interval is 1, the padding is 1, the data dimension is regularized using Batch Normal, and the activation function is ReLU. In addition, since fish features are more difficult to extract, the residual results are used for enhanced extraction, and finally the fish pattern spot category is output.

[0046] Through this step and step 1, the fish pattern spot category, its size spacing ratio and swimming speed can be predicted, and the direct prediction of the actual size of the fish is avoided; that is, three input features of fish pattern spot category, size spacing ratio and swimming speed are obtained.

[0047] Step 3: Construction of a temporal network; the feature vector of each fish in the target fish school is regarded as an input vector in the temporal network, where the feature vectors are: fish pattern spot category, size spacing ratio between fish and swimming speed of each fish; first, the multi-dimensional features of each fish can be compressed into a sequence vector. In addition, in an underwater environment, the behavior and arrangement of fish schools have temporal correlation. Using the features of continuous fish as the input vector of the temporal network can better capture the temporal relationship within the fish school, thereby more accurately analyzing and predicting the behavior and changes of the fish school. However, inputting more feature vectors of fish at the same time will greatly increase the computational complexity of the mobile device, and it is unnecessary for the recognition of static fish. Therefore, a temporal analysis network framework based on the swimming speed of fish is proposed, such as Figure 4 As shown in the figure, the first layer of the network is the Del-Act activation function, which determines the number of input vectors. The Del-Act activation function is:

[0048]

[0049] Where num represents the number of feature vectors to be input, which is a maximum of 10; v iRepresented as the speed value of the ith fish; the Del-Act activation function formula can be used to determine the number of input vectors according to the fish swimming speed, that is, the number of inputs can be dynamically adjusted according to the activity level and speed of the fish; this is because a faster swimming speed means that the relationship between individuals in the fish school changes faster; by using more input vectors, this temporal change can be better captured to more accurately analyze and predict the behavior of the fish school. In addition, selecting the number of input vectors according to the swimming speed of the fish can make the model dynamically adaptable; for fast-moving fish schools, increasing the number of inputs can provide more information, while for slower fish schools, reducing the number of inputs can save computing resources;

[0050] The number of input vectors is determined by the Del-Act activation function, and the subsequent network layer is a Random decision operator, which mainly

[0051] The function is to determine the starting point of the input vector. The Random operator is:

[0052] Random(1,8-num);

[0053] Among them, num is the number of feature vectors to be input, and Random(1,8-num) means randomly selecting a positive integer from 1 to 8-num, including 1 and 8-num;

[0054] The Random decision operator formula determines which vectors are input into the temporal network. For example, if num is 3 and the number randomly generated by the Random decision operator is 4, then the vectors input into the overall network are the feature vectors of the three fishes Fish-4, Fish-5, and Fish-6.

[0055] Finally, an adaptive amplification factor for static fish identification is proposed. When num is small, that is, when the fish presents static characteristics, it is given by analyzing the accuracy of the training of the specific fish pattern spot category. The adaptive amplification factor Factor is:

[0056]

[0057] Among them, Factor is the adaptive amplification factor, Arc(class) is the accuracy of fish pattern spot classification, and the entire network can only operate when the value of Arc(class) is greater than 0.9;

[0058] Furthermore, the network algorithm can be adjusted according to the accuracy of the classification of fish pattern spots through the formula of adaptive amplification factor Factor. When the accuracy of the classification of fish pattern spots is low, the Factor is large, and its value is calculated and rounded up. According to the integer obtained, the number of random draws by the Random decision operator is determined. For example, when Factor is 2 and num is 3, Random is chosen twice, 2 and 3 respectively; the first set of input vectors selected is Fish-2, Fish-3, Fish-4; the second set is Fish-3, Fish-4 and Fish-5. The final output result is added according to the output results of the two sets of input vectors to see which fish has the highest confidence. The basic network architecture in the present invention is LSTM, Batch Normal and Leaky ReLU.

[0059] Step 4: Ship movement and feed feeding. In step 3, the type of fish school was determined, that is, the type of feed that needs to be fed was determined. In this step, the sonar device on the unmanned ship is used to emit sound waves underwater to measure the time it takes for the sound waves to reach the fish and return from the ship, calculate the distance, and then the unmanned ship sails to the location of the fish school to feed the feed accurately.

[0060] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Any technician familiar with the profession, without departing from the scope of the technical solution of the present invention, according to the technical essence of the present invention, any simple modification, equivalent replacement and improvement made to the above embodiment still falls within the protection scope of the technical solution of the present invention.

Claims

1. A method for feeding a target fish school by an unmanned boat based on a temporal network algorithm, characterized in that: The specific steps include: Step 1: Preliminary collection of target fish data; specifically collect the swimming speed and position of the target fish; Step 2: Extracting target fish data features; The target fish data feature extraction is specifically as follows: feature extraction is performed on the target fish picture, and the extracted features include: fish pattern spot category and size spacing ratio between fish, wherein the fish pattern spot category is extracted using a network architecture based on YOLO transfer learning, the fish category is set to 1, and the remaining options are all 0; in addition, the regression frame detection part of the YOLO algorithm is intercepted, and a prediction frame is output, and fish spacing is extracted through the prediction frame and according to the size spacing ratio formula; The size spacing ratio formula is: Where Sp is the size spacing ratio between two adjacent fishes, Ce1 and Ce2 are the center points of the target detection frames of the two adjacent fishes, X2 and X1 are the upper right corner horizontal coordinate value and the upper left corner coordinate value of the target detection frame of the right fish, and X′2 and X1′ are the upper right corner horizontal coordinate value and the upper left corner coordinate value of the target detection frame of the left fish. Step 3: Construction of temporal network; The construction of the temporal network is specifically as follows: the feature vector of each fish in the target fish school is regarded as an input vector in the temporal network, wherein the feature vector is: the fish pattern spot category, the size spacing ratio between the fish and the swimming speed of each fish; the Del-Act activation function is used to determine the number of input vectors, the Random decision operator is used to determine the starting point of the input vector, and finally a static fish recognition adaptive amplification factor Factor is proposed to determine the number of extractions of the Random operator; The Del-Act activation function is: Where num represents the number of feature vectors to be input, which is a maximum of 10; v i Represented as the speed value of the i-th fish; The Random decision operator is: Random(1,8-num); Among them, num is the number of feature vectors to be input, and Random(1,8-num) means randomly selecting a positive integer from 1 to 8-num, including 1 and 8-num; The adaptive amplification factor Factor is: Among them, Factor is the adaptive amplification factor, Arc(class) is the accuracy of fish pattern spot classification, and the entire network can only operate when the value of Arc(class) is greater than 0.9; Step 4: Ship movement and feed feeding; The ship movement and feed feeding are specifically as follows: through the sonar device on the unmanned ship, sound waves are emitted underwater to measure the time it takes for the sound waves to reach the fish and return from the ship, calculate the distance, and then the unmanned ship sails to the location of the fish school to feed the feed accurately.

2. The method for feeding a target fish school by an unmanned boat based on a temporal network algorithm according to claim 1, characterized in that: The preliminary collection of target fish school data is specifically: collecting data for each fish in the target fish school, wherein the data collection object is ten consecutive fish in the target fish school, and specifically collecting the swimming speed of these ten consecutive fish and the position of the target fish school.

3. The method for feeding a target fish school by an unmanned boat based on a temporal network algorithm according to claim 2, characterized in that: The collection of the swimming speeds of ten consecutive fish in the target fish school is specifically as follows: the swimming speed of each fish in the target fish school is obtained by using a radar to measure the frequency and phase changes of the echo signal of the movement speeds of the ten consecutive fish in the target fish school.

4. The method for feeding a target fish school by an unmanned boat based on a temporal network algorithm according to claim 3, characterized in that: The acquisition of the position data of the target fish school is specifically as follows: using an underwater high-definition camera to collect images of the fish school positions scanned by the radar to obtain a picture of the target fish school.

Images