Design Of Shellfish Seafood Size Detection System Based On Deep Learning

The grading of shellfish seafood plays an important role in the sale of shellfish seafood,which can directly affect the sales price of shellfish seafood.Grading according to the specification and size can significantly increase the selling price and bring higher profits to the enterprise.Manual sorting and manual weighing have the problems of slow sorting speed,low precision and consuming a lot of manpower.The size detection of shellfish seafood is an important part of the visual sorting equipment for shellfish seafood.The detection usually uses traditional visual detection algorithm,which requires fine parameter setting,and has the disadvantages of low precision,slow speed and poor generalization.A good visual size detection algorithm is conducive to the classification of shellfish seafood.In view of the above problems,this paper designs a shellfish seafood size detection system based on deep learning.The main research contents are as follows:The size detection system of shellfish seafood is designed,the basic composition of the size detection system is analyzed,and the hardware scheme and software scheme of the size detection system are designed.According to the actual requirements of size detection,the evaluation indexes of the core algorithm for size detection are designed: m Io U,F1 score and FPS,in which m Io U is required to reach more than 95%,F1 score is not less than 99.5%,and FPS not less than 100.According to the detection requirements,the evaluation indexes of the system operation are designed: the relative error of area detection is required to be within ±5%,the system detection rate is higher than 10 FPS,and the system can run stably.A shellfish seafood dataset under the size detection scenario is constructed,including 900 images of three kinds of shellfish seafood: oyster,abalone and conch.In addition,the classification and pixel-level segmentation and annotation of the dataset are carried out,and the image preprocessing and image enhancement methods applicable to the dataset are designed.A shellfish seafood segmentation and classification method based on convolutional neural network encoding/decoding mode is proposed.In order to obtain the multi-scale semantic information of the image,the encoding module of the network adopts the parallel double branch network structure to enhance the deep semantic feature representation ability while preserving the spatial details of the image.In addition,the channel split and shuffle module and depth convolution module are used as the basic modules to improve the running speed of the algorithm.The decoding module of segmentation integrates multi-scale features to fully extract the context multi-scale information.In the experiment,through testing of 1200 data collected in the sorting site,the m Io U reaches 96.693%,the F1 score reaches 99.872%and the FPS reaches 108 FPS.The experimental results show that this method can quickly obtain the segmentation area and category of the target,and has high segmentation accuracy.On the Qt development platform,the shellfish seafood size detection software is developed using C++ language.The software can obtain the area,length,width and category of shellfish seafood.The design of user login interface,main interface and parameter setting interface are realized.The deep learning model is accelerated and deployed with Open VINO tool.Finally,the system can run stably at the speed of 15 FPS,and the average relative error of area detection is within ±5%.