Phenotypic traits such as body weight and body length of fish are very important economic traits in aquaculture and genetic breeding. In order to avoid the uncertainty, error randomness and low efficiency of manual measurement, this paper develops an automated, non-invasive device based on Mask Region Convolutional Neural Network (Mask R-CNN) for fish image segmentation and phenotypic traits measurement. The device consists of two parts: an image acquisition device able to measure fish of different sizes (body length 1-40 cm) and control software. The control software based on Mask R-CNN can train and predict the target traits of images, and realize the measurement, storage and management of target data. The experimental results show that the average relative error in body length and body height of Larimichthys crocea measured by the device is less than 4%. The body weight was fitted with multiple regression models based on body length, body height and body surface area. The correlation coefficient between measured values and the real body weight was 0.99, the average relative error was 4%, and the average processing time for each image was 3 seconds, which was 8 times as fast as manual measurement. The data measurement device based on machine vision and image capture developed in this study can automatically, efficiently and accurately obtain morphological and weight data of L. crocea, which provides a more convenient and efficient phenotype evaluation tool for the evaluation of L. crocea germplasm resources, breeding of improved varieties and germplasm innovation.