Study On Posture Detection Method And Adjustment Technology Of Tilapia Before Gutting

China is the largest tilapia breeding and exporting country in the world.In recent years,Chinese tilapia processing industry has developed rapidly.The initial processing of fish includes the removal of viscera and other processes.Due to the requirements of orientation of head,tail,abdomen and back and single feeding,gutting is the bottleneck of the processing of tilapia,which affects the processing rhythm of the whole production line.Presently,there is little application of fish orientation equipment in the aquatic processing enterprises.The main method is to adjust fish posture one by one and feed them into the gutting machine.With the increase of labor cost,it is urgent to develop fish posture detection and adjustment technology as well as equipments.The existing fish body posture detection methods have poor robustness,and the posture adjustment technology relies on single-line preprocessing,which is easy to block.Therefore,tilapia were used as the research object and a fish posture detection model based on deep learning was proposed,which realized the online adjustment of multi-target tilapia posture.A set of tilapia posture detection and adjustment system relying on machine vision technology has been developed,which provided technical support for the automation and intelligence of the tilapia primary processing production line.The main research contents and conclusions of this paper are as follows:1.A longside posture definition was proposed to quantitatively define the posture information of fishTo solve the problem of difficult quantification of fish posture information,a longside posture definition(LPD)was proposed,which defined fish posture as a combination of angle and direction value.The outer rectangle of the fish body region was placed in cartesian coordinates.The angle through which the longside of the rectangle rotates counterclockwise to the horizontal direction around the end point with the smallest ordinate value was set as fish angle.The rectangle was divided diagonally into four areas,representing the head,tail,abdomen and back respectively,and assigned binarization values.According to the relative position of fish head,tail,abdomen and back,encoding was converted to the unique direction value of the current posture.Accurate quantification of fish posture information was achieved by LPD,which laid a foundation for the subsequent deep learn-based fish posture detection.2.The method of fish posture detection was studied.The fish posture detection model based on deep learning was constructed to realize the real-time detection of fish posture and position information under multi-objectiveAiming at the poor robustness of traditional image processing techniques for fish posture detection,a fish posture detection method based on deep learning was studied.On the basis of YOLOv5 s model,the fish Angle value was discretized by circular smooth label(CSL).The difference between the predicted angle and the actual angle was calculated by Gaussian function as a window function,so that the model can learn the angle information.In order to solve the class imbalance of fish angles in the data set which caused by angle discretization and increase the diversity of fish images in the training process,off-line rotation enhancement and online noise enhancement algorithms were designed respectively to amplify the sample data.The prediction channel of angle value and direction value was designed.Finally,the YOLOv5 s model could detect the angle value,direction value and center position of the fish body in real time under the dynamic multi-objective.Through experiments,the CSL window radius of 2 and rotation enhancement scaling factor of 0.8 were the best model parameter combination.On the test set,the precision of the model was 98.33% and the AP50 was 0.902.The fish posture detection speed was 50 fps and 22.2 fps on GPU and CPU platform,respectively.3.The on-line adjustment technology of fish posture was studied.The fish body orientation device was constructed and tested for verification and optimizationCombined with the real-time position and posture information of fish detected above,the grasping performance of different end-effectors was compared and analyzed based on SCARA robot arm.Finally,the pneumatic flexible claw was selected as the end-effector of fish posture adjustment.Through experiments,the optimal grasping position of fish was 0.4 times the length of fish body from fish head.Fish grabing,head and tail adjustment and position movement was realized.A segmental beveled posture adjustment device for fish abdomen and back was designed based on the physical characteristics of fish.The influencing factors of abdomen and back adjustment were analyzed.Through orthogonal test,the optimal combination of parameters in the process of abdomen and back adjustment were as follows: the length of the lower beveled plane was 115 mm,the length of the upper beveled plane was 116 mm,the inclination of the lower beveled plane was 82.5°,the inclination of the upper beveled plane was 47.5°,the distance between the lower beveled plane was 80 mm,and the ratio of the distance between the falling point and the edge line of the upper beveled plane was 0.3.At this time,the success rate of abdomen and back adjustment was 99.5%.4.A fish posture detection and adjustment software was developed,and a fish posture detection and adjustment system based on machine vision was developedUsing Modbus TCP as the communication protocol,the software of fish posture detection and adjustment was developed based on QT,a cross-platform graphical user interface application development framework.Based on the above research,a fish posture detection and adjustment system was developed.The system test results show that the posture detection precision was97.56% in the process of fish transportation,reaching the highest detection accuracy level required by posture adjustment.At the current maximum operating speed of the robot arm,the adjustment speed of fish posture was 0.505 pieces/second,and the adjustment success rate was95.3%.The source of experimental error was analyzed: when the fish to be oriented on the conveyor belt were closely attached to each other,it would be difficult for the end-effector to separate the fish body.