THE ESTABLISHMENT AND APPLICATION OF A FAST PHENOTYPIC DETERMINATION TECHNIQUE BASED ON FASTER R-CNN FOR GROWTH TRAITS IN SHRIMP

High-throughput genotyping and phenotyping are the key techniques for efficient and precise breeding. Compared with genotyping technology, the development of high-throughput phenotyping is relatively slower. In breeding of aquatic animals, especially for shrimp, the phenotypic data of the external growth traits are mainly obtained by manual measurement, which needs high labor intensity with low efficiency. In recent years, the rapid development of deep learning has provided technical support for high-throughput phenotyping. In order to improve the efficiency of obtaining phenotypic data of external growth traits of shrimp, we applied a Faster R-CNN (Faster Region-convolutional neural networks) model based on Region Proposal Networks (RPN) to automatically identify the body length of shrimp. Based on the training to 8400 shrimp photos, the model could rapidly identify and output location information of shrimp, and the total lengths of shrimp were accurately measured. For the shrimp photos taken in vertical view, the length of the recognition frame was highly correlated with the manually measured full length of the shrimp. For the photos taken in side view, some individuals of shrimp were in bent shape which will affect the correlations. We found that the ratio of the diagonal length of the recognition frame to the length (K value) could represent the degree of bending of the shrimp. Further analysis illustrated that the shrimp is straight in the picture and the length of the recognition frame is highly correlated with the full length of the shrimp when the K is less than 1.04. However, most of the shrimp in the photo is bent, and the diagonal length of the recognition frame is highly correlated with the full length of the shrimp, when the K is greater than 1.04. Consequently, we established a high-throughput technique to determine the full-length of shrimp. The establishment of this technique can save time in comparison to manual measurements on the shrimp phenotype and improve the efficiency of genomic selection breeding of shrimp. In addition, the establishment of this model also provides a new idea for the determination on the other external phenotypic data of shrimp, such as cephalothorax length and body segment lengths in shrimp, and lays an important foundation for the establishment of phenomics data in shrimp.