Research On Measurement Of Phenotypic Data Of Macro Shrimp Based On Deep-learning Compressed Model

As the source of the industrial chain,aquatic seedlings are the foundation of the aquatic industry.The quality of aquatic seedlings determines the success of aquaculture and the lifeline of the future industry.Seed is known as the "chip" in the aquaculture industry,China’s seed industry must not only overcome the challenges of the natural environment,but also reduce its dependence on foreign imports.In response to the severe challenges of aquatic germplasm resources safety,many research institutes have begun to tackle key problems in seedling cultivation and selection.The phenotypic data of aquatic organisms is an important reference factor in the breeding and optimization process,but there is currently a lack of corresponding intelligent equipment in China.This thesis takes the deep learning compression model for phenotypic data determination of Macrobrachium as the research goal,hoping to reduce the complexity of the model,facilitate deployment on hardware platforms and promote the portability and intelligence of phenotype data determination equipment.This thesis takes the measurement of phenotypic data of Macrobrachium as the research goal,researches the model compression techniques such as knowledge distillation and model quantification and reduces the number of parameters of the deep neural network and improves the efficiency of the algorithm.The main contributions and innovations of the thesis are in the following aspects:1.The thesis propose a knowledge distillation method based on multi-layer fusion of isomorphism model,which combines the feature fusion method of mixed attention mechanism and the loss function based on Wasserstein distance.Multi-layer fusion features are applied to teacher-student models with large differences in complexity.This distillation method can maximize the ability of the teacher model to extract features so that the efficiency of knowledge distillation can be improved.2.A strategy of depth-based space narrowing and stepwise quantization finetuning is proposed and implement a mixed-precision quantization method based on architecture search.Experiments show that the two optimization methods can effectively speed up the speed of the model searching under the premise of ensuring the accuracy.3.A phenotypic data determination and classification system for Macrobrachium prawns was designed.Based on this system,the body length,second foot length and color information of Macrobrachium can be determined accurately,which were used to classify the Macrobrachium correctly and obtain the characteristics of the corresponding population.In thesis,the HRNet network is optimized by reducing the number of convolution kernels in the bottleneck structure and adding the distance-based constraints,so as to realize the detection of key points of Macrobrachium.