Intelligent feeding system for buoyant feed of large-mouth bass based on land-based circular pond breeding

Abstract

The application relates to an image data processing system applied to agriculture, and discloses an intelligent feeding system for floating feed of land-based round-pool cultured large-mouth bass, which comprises a visual acquisition module, an edge calculation processing module, an intelligent decision module and an intelligent feeding control module which are sequentially connected; the application proposes a special visual identification process (HSV+ morphology+ MOG2 background reduction) for floating feed, and then the operation parameters are systematically optimized through an orthogonal test, so that the system can reach a precision level that can be applied in actual breeding environment; the system has low deployment cost, low operation power consumption, low network dependence and high industrial applicability, and realizes real "on-site intelligence".

Description

Technical Field

[0001] This invention relates to an image data processing system for agriculture, and in particular to an intelligent feeding system for floating feed of largemouth bass in land-based circular ponds. Background Technology

[0002] Largemouth bass (Micropterus salmoides) is an important freshwater economic fish in my country, with national aquaculture production exceeding 980,000 tons in 2024. With the rapid development of recirculating aquaculture system (RAS) technology, land-based circular pond aquaculture has been widely used in intensive aquaculture due to its advantages such as water and land conservation, environmental control, and high stocking density.

[0003] Floating pellet feed has become the mainstream feed type for largemouth bass factory farming due to its advantages such as good suspension and ease of observation of fish feeding status. However, in actual production, feeding management still mainly relies on manual experience or timed mechanical feeding, which has the following prominent problems: overfeeding leads to feed waste and water quality deterioration, with generally high feed conversion ratios (1.5-2.0) and uneaten feed rates reaching 15%-25%; underfeeding affects fish growth performance and prolongs the farming cycle; manual feeding is labor-intensive and makes it difficult to achieve high-frequency, precise feeding.

[0004] In recent years, computer vision and Internet of Things technologies have been introduced into the field of aquaculture feeding. Existing research mainly includes: (1) Feeding behavior analysis method based on acoustic features: the feeding sound of fish is collected by hydrophone to judge appetite, but acoustic sensors are expensive and easily affected by environmental noise; (2) Feeding desire assessment model based on light-sound coupling: the system integrates visual and acoustic information for decision-making, but the system is complex and the deployment cost is high; (3) Behavior recognition system based on deep learning: the complex neural network model is run on GPU server, which has problems such as high power consumption (>300W), strong network dependence, and complex deployment, and is difficult to promote and apply in actual industrial aquaculture scenarios.

[0005] The market needs an integrated solution that is optimized for the specific conditions of land-based circular pond aquaculture environments (such as water surface reflection, bubble interference, and high-density fish aggregation) and can provide a low-cost, low-power, and easy-to-deploy solution based on edge computing. This invention addresses these issues. Summary of the Invention

[0006] To address the shortcomings of existing technologies, the present invention aims to provide an intelligent feeding system for floating feed of largemouth bass in land-based circular ponds. This system can be applied on a large scale in aquaculture sites, solves the problem of dynamic adjustment of feeding status in real time, and has high recognition accuracy and system stability.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A smart feeding system for floating feed of largemouth bass in land-based circular ponds includes: a visual acquisition module, an edge computing processing module, a smart decision-making module, and a smart feeding control module connected in sequence.

[0009] (1) Visual acquisition module: The industrial camera installed above the land-based circular pool accesses the camera video stream via the RTSP protocol to acquire continuous image frames and transmits the real-time video stream to the edge computing processing module;

[0010] (2) Supplemental lighting control module: Supplemental lighting equipment to maintain the brightness of the water surface;

[0011] (3) Edge computing processing module: The embedded edge computing device performs image preprocessing on each frame of the image; the candidate region of floating feed on the water surface is extracted based on the threshold segmentation of the HSV color space; the morphological opening and closing operations are used to eliminate water surface ripples and bubble noise; the background subtraction algorithm is used to separate dynamic feed particles from the static water surface background; the contour of feed residue is extracted by the contour detection algorithm and the area is filtered to calculate the number of feed residue N; when N is lower than the preset threshold, a feeding trigger signal is sent.

[0012] (4) Intelligent decision-making module: Generates feeding control instructions based on the output of the edge computing processing module, including a feeding strategy configuration unit and a real-time decision-making unit;

[0013] (5) Intelligent feeding control module: Receives control commands through the intelligent feeding control system, feeding execution module and feeding system device to execute precise feeding.

[0014] The aforementioned intelligent feeding system for floating feed of largemouth bass in land-based circular pond aquaculture specifically includes image preprocessing as follows: preprocessing each frame of the image, performing Gaussian filtering for noise reduction, and converting the color space from BGR to HSV.

[0015] The aforementioned intelligent feeding system for floating feed of largemouth bass in land-based circular pond aquaculture also includes an edge computing processing module: a parameter adaptive system that determines the optimal combination of operating parameters through orthogonal experimental methods.

[0016] In a further preferred embodiment, the optimal combination of operating parameters for the aforementioned intelligent feeding system for largemouth bass with floating feed based on land-based circular pond aquaculture includes: (1) Camera pixels: a 4-megapixel industrial camera with an image resolution of 2560x1440; (2) Supplemental lighting control: maintaining the water surface illuminance in the range of 8,500-10,000 Lux through LED supplemental lighting equipment; (3) Feed adaptation: floating pellet feed with a particle size of 2mm is preferred.

[0017] The aforementioned intelligent feeding system for largemouth bass with floating feed based on land-based circular pond aquaculture includes the following process:

[0018] (1) Initialization phase: After the system starts, the visual acquisition module begins to acquire video streams;

[0019] (2) Trial feeding stage: The first feeding lasts for 5-15 minutes to trigger the feeding behavior of the fish;

[0020] (3) Residual feed determination stage: The edge computing processing module continuously analyzes the video stream and continuously monitors and calculates the number of residual feed particles at a frequency of 10-20 minutes.

[0021] (4) Decision adjustment stage: The intelligent decision module dynamically adjusts the feeding decision based on real-time monitoring results, including whether to continue feeding;

[0022] (5) Termination judgment stage: When the number of residual particles exceeds the preset threshold, the current feeding is terminated;

[0023] (6) Cyclic waiting phase: After a preset waiting time, return to the trial feeding phase to start the next feeding cycle.

[0024] The aforementioned intelligent feeding system for floating feed of largemouth bass in a land-based circular pond includes: a pole, an industrial camera cantilevered on the pole to capture images of the floating feed area on the water surface, a feed hopper and an automatic feeder fixed on the pole to store floating pellet feed, and an edge computing control box fixed on the pole. The industrial camera is cantilevered on the pole. The edge computing control box contains a Raspberry Pi 5 single-board computer, relays, and other core control units.

[0025] In a further preferred embodiment, the aforementioned intelligent feeding system for floating feed of largemouth bass in a land-based circular pond includes a feeding execution module comprising a relay control unit and a feeder built into a control box.

[0026] In a further preferred embodiment, the aforementioned intelligent feeding system for floating feed of largemouth bass in a land-based circular pond aquaculture uses a Raspberry Pi 5 single-board computer with a floating feed particle recognition algorithm as the embedded edge computing device.

[0027] The advantages of this invention are:

[0028] Low cost, low power consumption, and easy deployment: This invention uses an embedded edge computing platform (such as Raspberry Pi 5) to replace the traditional PC-based GPU server, with a total power consumption of less than 15W and a reduction in equipment cost of over 80%. The edge deployment mode does not rely on cloud computing and network transmission, making it suitable for large-scale distributed deployment in aquaculture sites.

[0029] High recognition accuracy: Through multi-parameter collaborative optimization, with the optimal configuration of a 4-megapixel camera, 8,500-10,000 Lux illumination, and 2mm particle size feed, the system controls the false recognition rate of floating feed particles to within -3%, meeting the accuracy requirements of industrial precision feeding.

[0030] Closed-loop intelligent control: This invention constructs a complete closed-loop system of "visual recognition - uneaten bait analysis - strategy decision-making - precise feeding", realizing automated operation from perception to execution. It can dynamically adjust the feeding strategy according to the real-time feeding status of the fish, effectively avoiding overfeeding and underfeeding.

[0031] Significant economic benefits: After applying the system of this invention, it is expected to reduce the feed conversion ratio by 0.2-0.3, reduce the residual feed rate by 30%-50%, reduce the load on the circulating water treatment system by more than 20%, and improve the overall aquaculture efficiency by 15%-20%;

[0032] Excellent environmental adaptability: Through the supplemental lighting control module and parameter adaptive system, the system can operate stably under different weather and lighting conditions, adapting to various aquaculture environments such as direct sunlight on sunny days, diffused light on cloudy days, and weak light on overcast days. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the overall architecture of one embodiment of the intelligent feeding system of the present invention;

[0034] Figure 2 This is a schematic diagram of one embodiment of the intelligent feeding control system of the present invention;

[0035] Figure 3 This is a schematic diagram of one embodiment of the feeding system device of the present invention; Figure 4 This is the main effect diagram of the three factors of the present invention (light intensity effect diagram ( Figure 4 a), Camera pixel effect diagram ( Figure 4 b), Camera pixel effect diagram ( Figure 4 b)))

[0036] Figure 5 This invention relates to the interaction effect (interaction mode) between light intensity and feed particle size. Figure 5 a) The interaction effect between light intensity and camera pixels (Figure 5b));

[0037] Meaning of the reference numerals in the diagram:

[0038] 1 Industrial camera, 2 Pole, 3 Feed hopper, 4 Automatic feeder, 5 Edge computing control box. Detailed Implementation

[0039] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0040] like Figure 1 As shown, a smart feeding system for floating feed of largemouth bass in a land-based circular pond includes: a visual acquisition module, an edge computing processing module, a smart decision-making module, and a smart feeding control module connected in sequence.

[0041] 1. Visual Acquisition Module: The module acquires continuous image frames by accessing the camera video stream via the RTSP protocol through an industrial camera installed above the land-based circular pool, and transmits the real-time video stream to the edge computing processing module.

[0042] As a preferred embodiment, the industrial camera is an RTSP 1080P camera.

[0043] Second, supplemental lighting control module: supplemental lighting equipment to maintain the brightness of the water surface;

[0044] III. Edge Computing Processing Module:

[0045] 3.1 Image preprocessing is performed on each frame of the image using an embedded edge computing device; as one embodiment, the image preprocessing specifically includes preprocessing each frame of the image, performing Gaussian filtering for noise reduction and BGR to HSV color space conversion. As a preferred embodiment, the embedded edge computing device uses a Raspberry Pi 5 single-board computer deployed with a floating feed particle recognition algorithm.

[0046] The specific process is as follows:

[0047] Gaussian filtering and BGR to HSV color space conversion are performed on each frame of the image.

[0048] (1) Gaussian filtering noise reduction

[0049] The image is smoothed using a 5×5 Gaussian kernel convolution. The two-dimensional Gaussian function is:

[0050]

[0051] The filtered image is:

[0052]

[0053] As a preferred embodiment, the standard deviation σ is set to 1.4, and the kernel weights are normalized to eliminate high-frequency noise from water surface reflection and sensor imaging noise.

[0054] (2) BGR to HSV color space conversion

[0055] Each pixel of the filtered image is transformed from BGR space to HSV space. Let R, G, B ∈ [0, 255], and denote:

[0056]

[0057] The calculations for each component are as follows.

[0058] Brightness V:

[0059]

[0060] Saturation S:

[0061]

[0062] Hue H:

[0063]

[0064]

[0065] After conversion, H∈[0,360], S∈[0,255], and V∈[0,255]. As a preferred embodiment, H is compressed and mapped to [0,180] to adapt to the 8-bit integer storage of the OpenCV computation library. The floating feed particles are brownish-yellow, and their hue is highly distinguishable from the blue-green water background in HSV space; therefore, HSV space is used for subsequent segmentation.

[0066] 3.2 Extracting candidate regions for floating feed on the water surface based on HSV color space threshold segmentation;

[0067] The converted HSV image is subjected to a three-channel joint thresholding process pixel by pixel to generate a binary mask image M(x,y):

[0068]

[0069] Pixels with a mask value of 255 constitute the candidate region for floating feed. As a preferred embodiment, for brownish-yellow floating feed particles, H∈[8,35], S∈[80,255], and V∈[60,255] are selected; the above thresholds can be corrected based on the calibration images before and after the supplementary lighting control module is turned on.

[0070] 3.3 Morphological opening and closing operations are used to eliminate water surface ripples and bubble noise;

[0071] Define structural element B. As a preferred embodiment, a 3×3 elliptical structural element is used. The erosion and dilation operations are as follows:

[0072]

[0073] Where A is the binary image to be processed, B_z is the set of structuring elements translated to point z, and the structuring element in the dilation formula is its reflection set.

[0074] Opening operation (corrosion followed by dilation) is used to eliminate bubble highlights and fine ripple noise with an area smaller than the structural element.

[0075]

[0076] Closing operations (expansion followed by corrosion) are used to fill internal pores in feed pellets and connect areas of the same pellet that have broken due to corrugation.

[0077]

[0078] Perform one opening operation and one closing operation on the mask M in sequence to obtain the denoised mask image:

[0079]

[0080] 3.4 A background subtraction algorithm is used to separate dynamic feed particles from the static water surface background;

[0081] The Gaussian-filtered image is converted to grayscale and then float-point. The water surface background model B_t(x,y) is updated online using the moving average method.

[0082]

[0083] In the formula, I_t is the grayscale image of the t-th frame, α is the background learning rate, and the background model is initialized with the first frame. As a preferred embodiment, α=0.05, so that the background model can track the slow changes in water surface illumination without absorbing feed particles that stay there for a short time.

[0084] Calculate and binarize the difference between the current frame and the background model to extract the moving foreground:

[0085]

[0086] As a preferred embodiment, the differential threshold T_d = 25.

[0087] By taking the bitwise intersection of the motion foreground mask and the color segmentation result, an effective residual feed mask that simultaneously satisfies both "feed color characteristics" and "motion characteristics" is obtained, eliminating interference such as pool wall reflections and static floating objects on the water surface.

[0088]

[0089] 3.5 Extract the outline of the uneaten feed using a contour detection algorithm and perform area filtering to calculate the number of uneaten feed N; when N is lower than a preset threshold, send a feeding trigger signal. In a preferred embodiment, the preset threshold is 50.

[0090] (1) Contour extraction

[0091] A contour detection algorithm based on boundary tracking (Suzuki algorithm) is used to extract the outer contours of all connected regions of the binary mask M_2, resulting in a contour set {C_1, C_2, …, C_K}.

[0092] (2) Calculation of outline area

[0093] For each contour C_i, calculate its pixel area using Green's formula based on the contour vertex coordinates (x_j, y_j) (a total of n_i vertices):

[0094]

[0095] It is stipulated that the (n_i+1)th vertex coincides with the first vertex, i.e., (x_{n_i+1}, y_{n_i+1}) = (x_1, y_1).

[0096] (3) Area Filtering

[0097] Let A_p be the reference area of ​​a single floating feed particle on the imaging plane, and d be the imaging diameter of the feed particle (in pixels). Then:

[0098]

[0099] As a preferred embodiment, based on the camera installation height and 1080P resolution, d=12 pixels, corresponding to A_p≈113 pixels². The contour is filtered using [k_1·A_p, k_2·A_p] as the effective area range to obtain the effective bait contour set:

[0100]

[0101] In a preferred embodiment, k_1=0.3 and k_2=4.0; areas smaller than the lower limit are judged as residual noise, and areas larger than the upper limit are judged as floating debris or large-area reflections on the water surface, and are all rejected.

[0102] 3.6 Parameter adaptive system for determining the optimal combination of operating parameters through orthogonal experimental methods.

[0103] To improve the accuracy of visual judgment in feed pellet identification, this study conducted a three-factor orthogonal experiment, focusing on three factors: different light intensities, camera pixels, and feed pellet size. The experiment quantified the impact of these factors on the false judgment rate of pellet identification, analyzed methods to reduce the false judgment rate, and determined the optimal parameter combination. Light intensity was determined using a combination of natural light and LED supplementary lighting, categorized as low (<20,000 Lux), medium (20,000-40,000 Lux), and high (>40,000 Lux). Three network cameras with different pixel counts were selected as image acquisition devices: LeCheng TF1c (1.3 megapixels), Dahua P20A2W (2 megapixels), and Dahua P40A2WT (4 megapixels). The experimental feed consisted of commercial floating pellet feed, including Tongwei 8915 (2mm), Minghui 993 (3mm), and Tianma JZY01 (4mm). L27(3 3 The orthogonal experimental design was used, and the selected factor levels and orthogonal experimental scheme are shown in Table 1. The dependent variable is the absolute value of the misclassification rate, defined as |misclassification rate| = |(number of identified particles - number of actual particles) / number of actual particles| x 100%.

[0104] Table 1. Factor Level Table for Orthogonal Experiment

[0105]

[0106] Three-way ANOVA was used to assess the impact of the main effects of each factor and the interaction effects of two-factor and three-factor factors on the absolute value of the false positive rate. The statistical model is as follows:

[0107]

[0108] Where Yijk is the absolute value of the misclassification rate, μ is the population mean, Ai, Bj, and Ck are the three-factor main effects, (AB)ij, (AC)ik, and (BC)jk are the second-order interaction effects, and (ABC)ijk is the third-order interaction effect. The error is random. The F-statistic was calculated using Type II sum of squares, with a significance level set at p < 0.05. All analyses were performed using the Python 3.12 statsmodels library.

[0109] The analysis of variance results (Table 1) show that, with the absolute value of the false positive rate as the dependent variable, light intensity is significant (p<0.01) and is the primary factor affecting the magnitude of feed particle recognition error. The interaction effects between light intensity and camera pixels (A × B) and between light intensity and feed particle size (A × C) are also significant (p<0.05). In particular, the third-order interaction effect A×B×C is highly significant (p<0.001), indicating a complex synergistic mechanism among the three factors.

[0110] Table 2 Results of orthogonal experiment variance analysis

[0111]

[0112] The results of the orthogonal experiment (Table 3) show that when the absolute value of the error rate is used as the evaluation index, the importance of each factor is ranked in order of range R as follows: light intensity (A) > feed particle size (C) > camera pixels (B).

[0113] Table 3. Main Effect Analysis Results of Orthogonal Experiment

[0114] 1 A1 B3 C1 4.17 2 A1 B3 C2 5.31 3 A1 B3 C3 11.41 4 A1 B2 C1 6.65 5 A1 B2 C2 19.10 6 A1 B2 C3 7.46 7 A1 B1 C1 2.13 8 A1 B1 C2 6.17 9 A1 B1 C3 13.97 10 A2 B3 C1 11.02 11 A2 B3 C2 11.51 12 A2 B3 C3 9.46 13 A2 B2 C1 11.46 14 A2 B2 C2 8.81 15 A2 B2 C3 15.61 16 A2 B1 C1 15.81 17 A2 B1 C2 14.84 18 A2 B1 C3 5.77 19 A3 B3 C1 9.35 20 A3 B3 C2 17.23 21 A3 B3 C3 15.38 22 A3 B2 C1 12.97 23 A3 B2 C2 6.20 24 A3 B2 C3 7.32 25 A3 B1 C1 11.40 26 A3 B1 C2 9.87 27 A3 B1 C3 21.74 k1 8.49 11.30 9.44 k2 11.59 10.62 11.00 k3 12.38 10.54 12.01 Range R 3.90 0.76 2.57 Sort 1 3 2

[0115] Based on the results of the orthogonal experiment analysis, main effect plots of each factor on the absolute value of the error rate were created. The main effect plots of each factor are shown below. Figure 4 As shown. Regarding light intensity ( Figure 4 (a) The error rate was lowest under low light conditions, intermediate under medium light conditions, and highest under high light conditions, with a range of 3.90 percentage points. This result indicates that the error range of feed particle recognition increases with increasing light intensity. Regarding feed particle size ( Figure 4 c) The error rate is lowest at 2mm particle size, intermediate at 3mm, and highest at 4mm, showing a trend of increasing error with increasing particle size. Regarding camera pixels ( Figure 4 (b) The error amplitudes of the three pixel levels are very small, indicating that under the conditions of this experiment, the independent influence of camera pixels on the error amplitude is weak.

[0116] The interaction between light intensity and feed particle size is significant, and the interaction pattern is as follows: Figure 5 As shown in Figure 5a, under low light conditions, the error amplitude of 2mm particle size was the lowest, significantly better than that of 3mm and 4mm particle sizes; however, under high light conditions, the error amplitude of 4mm particle size increased sharply to approximately 14.8%, becoming the case with the largest error among all combinations. This cross-effect indicates that small-particle-size feed has stronger recognition stability under low light conditions, while the recognition accuracy of large-particle-size feed deteriorates sharply under high light conditions. The interaction effect between light intensity and camera pixels is also significant (Figure 5b), reflecting the differences in adaptability of different pixel devices under different lighting conditions.

[0117] The effects of light intensity, camera pixels, and feed particle size on the false recognition rate of feed particles were evaluated, and the importance was ranked as follows: light intensity > feed particle size > camera pixels. The optimal combination of operating parameters included: (1) camera pixels: a 4-megapixel industrial camera with an image resolution of 2560x1440; (2) supplemental lighting control: maintaining the water surface illuminance in the range of 8,500-10,000 Lux through LED supplemental lighting equipment; (3) feed adaptation: floating pellet feed with a particle size of 2 mm was preferred.

[0118] Fourth, intelligent decision-making module: generates feeding control instructions based on the output of the edge computing processing module, including a feeding strategy configuration unit and a real-time decision-making unit;

[0119] The number of uneaten bait N is calculated by area based on the sticky outline.

[0120]

[0121] For each valid contour, the ratio of its area to the reference area of ​​a single particle is rounded to calculate the particle count, which is then summed. N is then compared with a preset threshold N_0.

[0122]

[0123] As a preferred embodiment, a preset threshold N_0=50 is set. When there are less than 50 uneaten bait pellets on the water surface, it is determined that the feed has been fully consumed and the fish are actively feeding. A feeding trigger signal is then sent to the intelligent decision-making module. Conversely, if there are more uneaten bait pellets, feeding is suspended to avoid feed waste and water pollution.

[0124] Fifth, the intelligent feeding control module: This module receives control commands from the intelligent feeding control system, feeding execution module, and feeding system device to execute precise feeding. As one embodiment, the feeding execution module includes a relay control unit and a feeder built into the control box.

[0125] like Figure 2 As shown, the process of the intelligent feeding control system includes:

[0126] (1) Initialization phase: After the system starts, the visual acquisition module begins to acquire video streams;

[0127] (2) Trial feeding stage: The first feeding lasts for 5-15 minutes to trigger the feeding behavior of the fish;

[0128] (3) Residual feed determination stage: The edge computing processing module continuously analyzes the video stream and continuously monitors and calculates the number of residual feed particles at a frequency of 10-20 minutes.

[0129] (4) Decision adjustment stage: The intelligent decision module dynamically adjusts the feeding decision based on real-time monitoring results, including whether to continue feeding;

[0130] (5) Termination judgment stage: When the number of residual particles exceeds the preset threshold, the current feeding is terminated;

[0131] (6) Cyclic waiting phase: After a preset waiting time, return to the trial feeding phase to start the next feeding cycle.

[0132] like Figure 3 As shown, the feeding system includes: a pole, an industrial camera cantilevered on the pole to capture images of the floating feed area on the water surface, a hopper fixed on the pole to store the floating pellet feed and an automatic feeder, and an edge computing control box fixed on the pole. The industrial camera is cantilevered on the pole. The edge computing control box contains a Raspberry Pi 5 single-board computer, relays and other core control units.

[0133] This invention replaces traditional PC-based GPU servers with embedded edge computing, resulting in low deployment costs, low power consumption, minimal network dependence, and high industrial applicability, achieving true "on-site intelligence"—a fundamental difference from existing deep learning solutions. Furthermore, this invention proposes a dedicated visual recognition process (HSV + morphology + MOG2 background subtraction) for floating feed, rather than a general target detection algorithm, a core difference from existing vision solutions. Finally, this invention systematically optimizes operating parameters through orthogonal experiments, enabling the system to achieve a practical accuracy level in real-world aquaculture environments—a crucial leap from laboratory research to industrial application.

[0134] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the above embodiments do not limit the present invention in any way, and all technical solutions obtained by equivalent substitution or equivalent transformation fall within the protection scope of the present invention.

Claims

1. A smart feeding system for floating feed of largemouth bass in land-based circular ponds, characterized in that, include: The following modules are connected in sequence: visual acquisition module, edge computing processing module, intelligent decision-making module, and intelligent feeding control module. (1) Visual acquisition module: The industrial camera installed above the land-based circular pool accesses the camera video stream via the RTSP protocol to acquire continuous image frames and transmits the real-time video stream to the edge computing processing module; (2) Supplemental lighting control module: Supplemental lighting equipment to maintain the brightness of the water surface; (3) Edge computing processing module: Uses embedded edge computing devices to preprocess each frame of the image; Candidate regions for floating feed on the water surface are extracted based on HSV color space threshold segmentation; morphological opening and closing operations are used to eliminate water surface ripples and bubble noise; a background subtraction algorithm is used to separate dynamic feed particles from the static water surface background; the contour of uneaten feed is extracted by contour detection algorithm and area filtering is performed to calculate the number of uneaten feed N; when N is lower than a preset threshold, a feeding trigger signal is sent. (4) Intelligent decision-making module: Generates feeding control instructions based on the output of the edge computing processing module, including a feeding strategy configuration unit and a real-time decision-making unit; (5) Intelligent feeding control module: Receives control commands through the intelligent feeding control system, feeding execution module and feeding system device to execute precise feeding.

2. The intelligent feeding system for floating feed of largemouth bass in land-based circular ponds according to claim 1, characterized in that, The image preprocessing specifically involves preprocessing each frame of the image, including Gaussian filtering for noise reduction and BGR to HSV color space conversion.

3. The intelligent feeding system for floating feed of largemouth bass in a land-based circular pond as described in claim 1, characterized in that, The edge computing processing module also includes a parameter adaptive system that determines the optimal combination of operating parameters through orthogonal experimental methods.

4. The intelligent feeding system for floating feed of largemouth bass in land-based circular ponds according to claim 3, characterized in that, The optimal combination of operating parameters includes: (1) Camera pixels: a 4-megapixel industrial camera with an image resolution of 2560x1440; (2) Lighting control: maintaining the water surface illuminance in the range of 8,500-10,000 Lux through LED lighting equipment; (3) Feed matching: floating pellet feed with a particle size of 2mm is preferred.

5. The intelligent feeding system for floating feed of largemouth bass in land-based circular ponds according to claim 1, characterized in that, The intelligent feeding control system process includes: (1) Initialization phase: After the system starts, the visual acquisition module begins to acquire video streams; (2) Trial feeding stage: The first feeding lasts for 5-15 minutes to trigger the feeding behavior of the fish; (3) Residual feed determination stage: The edge computing processing module continuously analyzes the video stream and continuously monitors and calculates the number of residual feed particles at a frequency of 10-20 minutes. (4) Decision adjustment stage: The intelligent decision module dynamically adjusts the feeding decision based on real-time monitoring results, including whether to continue feeding; (5) Termination judgment stage: When the number of residual particles exceeds the preset threshold, the current feeding is terminated; (6) Cyclic waiting phase: After a preset waiting time, return to the trial feeding phase to start the next feeding cycle.

6. The intelligent feeding system for floating feed of largemouth bass in land-based circular ponds according to claim 1, characterized in that, The feeding system includes: a pole, an industrial camera cantilevered on the pole to capture images of the floating feed area on the water surface, a hopper and an automatic feeder fixed on the pole to store the floating pellet feed, and an edge computing control box fixed on the pole. The industrial camera is cantilevered on the pole. The control box is installed in the middle of the shelf and contains a core control unit such as a Raspberry Pi 5 single-board computer and relays.

7. The intelligent feeding system for floating feed of largemouth bass in a land-based circular pond as described in claim 1, characterized in that, The feeding execution module includes a relay control unit and a feeder built into the control box.

8. The intelligent feeding system for floating feed of largemouth bass in a land-based circular pond as described in claim 1, characterized in that, The embedded edge computing device uses a Raspberry Pi 5 single-board computer with a floating feed particle recognition algorithm.

Images