Mask_LaC R-CNN for measuring morphological features of fish. (15th November 2022)

• Record Type: Journal Article
• Authors: Han, Bing; Hu, Zhuhua; Su, Zhengwei; Bai, Xueru; Yin, Shuzhuang; Luo, Jian; Zhao, Yaochi
• Is Part Of: Measurement. Volume 203(2022)
• Publication Date: 2022-11-15
• Subjects: Morphological characteristics of fish -- Mask_LaC R-CNN -- Data expansion method -- Precision agriculture -- Aquaculture; Weights and measures -- Periodicals; Measurement -- Periodicals; Measurement; Weights and measures; Periodicals; 530.8
• DOI: 10.1016/j.measurement.2022.111859
• Languages: English
• ISSNs: 0263-2241

Abstract

Abstract: In aquaculture management, the measurement of various morphological characteristics of fish, such as body length, body width, caudal peduncle length and width, pupil diameter and iris diameter, are the main information basis for breeders to feed, use drugs, catch, grade and make breeding characters analysis. Obtaining these information quickly and accurately can provide effective guidance for management and control in the aquaculture process, and it is conducive to improve production efficiency and increase income. By combining data expansion method and Mask_LaC R-CNN network, this paper realizes the accurate measurement of various morphological characteristics of fish, and design a non-contact measurement system of fish morphological parameters. Because the fish position and image contrast in the data set are too single, data expansion method such as shrinkage transformation, translation transformation, contrast transformation and adding noise are used to simulate a more real scene. For the loss function of Mask R-CNN, smooth L1 loss is improved by using the balanced L1 loss function; the convolution decomposition is used to realize the lightweight network structure, so that the model can accelerate the network while maintaining high accuracy. At the same time, the hole convolution can effectively improve the accuracy of the detection algorithm without additional parameters and computational cost. The enhanced data set and the Mask_LaC R-CNN experiment shows that the improved scheme improves the measurement accuracy. Under the pure background, the mIoU of 50 test images is 0.930, the relative error of body length is 1.46%, and the relative error of body width is 0.65%. Under the complex background, the mIoU of 50 test images is 0.934, the relative error of body length is 6.98%, and the relative error of body width is 8.05%. Highlights: The loss function is improved. The designed lightweight structure and the hole convolution are used to improve the training speed and the detection accuracy, respectively. The training set is expanded. The experimental results show that the training model with data enhancement has higher accuracy in the measurement results. A non-contact measurement system of fish morphological parameters is proposed to realize the end-to-end measurement of fish morphological characteristic parameters.