Design And Experiment Of Mackerel Orientation Device Based On Machine Vision

The development of fish pre-treatment processing technology has a direct impact on the modernisation process of the fishing industry.As an important process before pre-treatment processing,fish orientation was currently mainly done manually,with problems such as noisy working environment,high labour intensity and low automation.In this paper,we had studied the traditional methods of fish orientation and found that the current fish orientation and finishing devices were mainly based on the physical characteristics of the head and tail of the fish,the different positions of the centre of gravity on the belly and back of the fish,and the differences in the friction coefficient and hardness of each part of the fish.In this paper,we propose an image processing and deep learning method to identify the orientation of the head,tail and back of the fish,and an actuator to orient the head,tail and back of the fish.A mackerel head,tail and ventral dorsal orientation test device has been constructed and its performance has been verified.The main findings of the study are as follows.(1)Design of the automatic orientation and finishing device for mackerel.The device consists of an image acquisition device,a fish b separation device,a fish belly and back orientation chute,a hardware control system and an image processing system for the host computer.The image acquisition device captures the fish on the conveyor belt and transmits the images to the host computer;the fish separation device separates the stacked fish,the hardware control system controls the movement of the cylinder pusher and the image processing system of the host computer identifies the head,tail and ventral and dorsal orientation of the mackerel.(2)Recognition of the ventral and dorsal orientation of fish based on image processing.Aiming at the obvious color difference between the belly and back of the mackerel,image processing technology was used to perform grayscale,denoising,binarization,contour extraction and other processing on the fish image.The average value of gray value of 3 points in the upper and lower parts of the fish body is extracted and compared,so as to identify the ventral and dorsal orientation of the fish.The verification test results show that the accuracy of this method is 100%.(3)Fish head and tail orientation recognition based on deep learning.In this paper,after collecting a complete fish image,the image was divided into "head" and "tail" to construct a dataset for fish orientation recognition.Four classical convolutional neural network classification models,including Alexnet,VGG16,Googlenet,and Res Net-18,were used for training.In order to prevent the model from overfitting due to the small sample size of the dataset,the model was trained by introducing transfer learning and improving the loss function.optimization.The experimental results showed that among the four network models,Res Net-18 has the best recognition performance on the test set.Its average recognition accuracy is 99.5%,the average recognition time is 0.034 s,and the F1 score is 99.49%.The network weights file size is moderate.(4)A directional arrangement device for the head,tail and abdomen and back of the mackerel was built.Taking mackerel as the test object,the performance verification test of fish head,tail and ventral dorsal orientation was carried out.The results show that the recognition accuracy of fish ventral and dorsal orientation based on image processing is 100%.The recognition success rate of the algorithm is 96.25%,the recognition time of the algorithm is 0.038 s,the average orientation time of each fish is3.725 s,and the fish orientation efficiency is about 15-16 fish/min.