| With the development of our economy,the people’s demand for seafood is also increasing.Squid is a high quality source of protein and is an important part of our pelagic fisheries.To meet the growing market demand,the squid fishing industry has to improve its production efficiency.Currently,there is still a huge proportion of manual labour in the squid harvesting process,which not only affects the overall efficiency and reduces the freshness of the squid,but also suffers from a lack of grading capabilities on some of the fishing vessels and a lack of added value to the labour.This creates a pressing need for a technology to help mechanise and intelligently upgrade the sorting chain,and with the continuous development of deep learning in recent years,there is some practical value in combining deep learning for application in the squid sorting chain.This thesis focuses on the application of shipboardbased deep learning target detection models to the squid sorting process as follows:Firstly,considering the low level of mechanisation in the current squid sorting process,a general solution design is proposed for the squid sorting line,with mechanical structures and control solutions based on the processes of loading,directional conveying,image acquisition and processing,and grading and sorting in the sorting process,to provide an environmental basis for the subsequent introduction of image sorting and processing technology solutions.Then,according to the characteristics of squid itself and the environment on board,the non-contact image processing is decided as the sorting and grading scheme.Through the measurement of various dimensions of squid,the high correlation between squid carcass length and body weight was confirmed,and the carcass length was established as the basis for grading.The traditional digital image processing method used in advance can better separate squid and background,but it is difficult to distinguish the carcass and head of squid and cannot detect the size of the carcass.To solve this problem,YOLO v5 target detection model with deep learning ability is introduced.By collecting and expanding squid images,data annotation of the images was performed to produce a squid dataset,which was put into the model for training.The model was improved according to the detection requirements.The performance of the improved model was improved compared to the previous one,and its AP value reached 93.8%.In preliminary tests,the effectiveness of the solution was demonstrated by means of offline image detection and real-time video detection.Finally,in order to further verify the accuracy of the model and improve the convenience of information interaction,an interactive graphical interface is designed and developed,which is combined with the detection model to form a program.Through the program,the sample squid was tested for multiple rounds.The experimental results showed that the average error rate of multiple rounds was 1.55%,which verified the accuracy of the program.Compared with the traditional manual sorting method,the sorting technology proposed in this thesis has the characteristics of accurate measurement,no contact and strong versatility. |
