Although genomic selection (GS) is accelerating the genetic improvement of economically important traits in aquaculture species, the accurate phenotyping of thousands of individuals remains a bottleneck in GS programs. Computer vision (CV) technologies with the assistance of deep learning may provide a cost-efficient solution for automated phenotyping of fish body size. In this study, we designed a CV-based automated phenotyping system based on the deep learning model Mask R-CNN (region-based convolutional neural networks) and tested its ability to predict standard length (SL) of tiger pufferfish (Takifugu rubripes) fed low fishmeal (LFM) diet. Simultaneously, we investigated the impact of the automated phenotyping on genetic parameter estimation and genomic prediction for growth-related traits. We raised a test population (n = 1001) that was fed the LFM diet from three months old until harvest at 22 months. Manual phenotyping of the fish was performed at 10-, 14-, 17-, and 22-months-old. At the last two measurement times, images of each fish were collected to train Mask R-CNN for fish detection and pixel-level segmentation of the fish body excluding the tail fin. An automated phenotyping pipeline combining the trained Mask R-CNN and standard CV algorithms were designed to estimate SL. We found a high correlation (0.965–0.971) and low relative difference (0.015–0.027) between the CV-derived and manually measured SL, indicating high precision of our CV system. A high correlation (0.912–0.956) was also observed between CV-derived body area and manual body weight. The heritability estimations and genomic predictions on manual SLs were conducted using a Genomic Best Linear Unbiased Prediction (GBLUP) model using a total of 16,471 single nucleotide polymorphism (SNP). Estimated heritability at the four sampling points ranged from 0.573 to 0.622, and the predictive ability of genomic prediction for harvest size was 0.627. These values are comparable to those reported in previous studies collected from the population fed a normal fishmeal diet. Our heritability estimation and genomic prediction derived from automated SL measurements differed negligibly to those based on manual measurements. This finding indicates the utility of our CV-based automated phenotyping system for GS breeding programs.