Automated Artemia detection and length measurement using deep convolutional networks? Gang Wang1[0000−0002−1916−6110] and Bernard De Baets1[0000−0002−3876−620X] KERMIT, Department of Data Analysis and Mathematical Modelling, Ghent University, Coupure links 653, B-9000 Gent, Belgium {Gang.Wang,Bernard.DeBaets}@UGent.be Automated image analysis has attracted much attention in the bioscience engineering field, due to its superiority in processing efficiency and assessment objectivity over manual image analysis [10]. In aquaculture, the brine shrimp Artemia is an intensively used cost-effective diet for fish and crustacean lar- vae, and recently, the number of Artemia studies is increasing. Artemia objects are usually observed by a stereo-microscope, and conventionally, Artemia im- age analysis tasks are carried out manually, which is rather time-consuming and labor-intensive. It is highly desired to have tailor-made methods for analyzing Artemia images. In response to such demands, we focus on two typical Artemia image analysis tasks: Artemia detection and Artemia length measurement. In many Artemia studies, e.g., in a quality assessment of Artemia hatching, an automated method for detecting and counting the Artemia objects in images would be highly desired, but few works have addressed this challenging task. Re- cently, deep convolutional neural networks (CNN) have been adopted for object detection [3]. A representative method is the one combining a region proposal module with a deep convolutional neural network (R-CNN method) [2]. But this method entails a considerable redundancy [4], since it uses anchor boxes to generate thousands of region proposals. Inspired by the R-CNN method, we pro- pose a so-called Marker-CNN method by combining a target marker proposal network with a CNN-based classifier [8], as illustrated in Fig. 1. In the first stage, instead of generating region proposals by blind anchor boxes, we design a target marker proposal network using a U-shaped fully convolutional network (UNet) architecture [5]. This module can indicate target candidates, separate adjacent objects and obtain the object structural information simultaneously. In the second stage, using a CNN architecture, we design a classifier to classify the target candidates into categories or label as a non-target, thereby obtaining the Artemia detection results. Furthermore, we have collected a dataset to train and test the proposed method. Experimental results on test images, which contain 1336 cysts and 3335 nauplii in total, manifest that the Marker-CNN method can accurately detect and count the Artemia objects in images, obtaining a detection accuracy of 99.6%. Samples of detection results are shown in Figs. 2(a) and 2(b). Artemia length is considered a key dependent variable in many Artemia stud- ies [1]. For instance, in a controlled pond Artemia production, the Artemia length ? Copyright c 2019 for this paper by its authors. Use permitted under Creative Com- mons License Attribution 4.0 International (CC BY 4.0). 2 G. Wang, B. De Baets Fig. 1. Illustration of the Marker-CNN method for Artemia detection. is usually adopted as a metric for evaluating the feeding strategies of intensive Artemia culture [6]. However, automated Artemia length measurement has so far not been addressed. Moreover, conventional image-based length measurement approaches cannot be readily transferred to measure the Artemia length, due to the distortion of non-rigid bodies, the variation over growth stages and the interference from the antennae and other appendages. To address this problem, we propose an automated Artemia length measurement method [9] using UNet and second-order anisotropic Gaussian kernels [7]. For a given Artemia image, the designed UNet model is used to extract a length measuring line structure, and, subsequently, the second-order Gaussian kernels are employed to transform the length measuring line structure into a thin measuring line with minimal loss of length measure. We have collected a dataset containing 250 training images and 100 test images. Experimental results show that our method can accurately measure the length of Artemia objects in images, obtaining a mean absolute percentage error of 1.16%. Sample results of length measuring line extraction obtained by our method are displayed in Figs. 2(c) and 2(d). (a) (b) (c) (d) Fig. 2. Samples of detection results (in blue/green rectangles) obtained by the Marker- CNN method (a-b), and samples of length measuring line extraction results (in red) obtained by our method (c-d). Automated Artemia detection and length measurement using CNN 3 References 1. Balachandar, S., Rajaram, R.: Influence of different diets on the growth, survival, fecundity and proximate composition of brine shrimp Artemia franciscana (kellog, 1906). Aquaculture Research 50(2), 376–389 (2019) 2. Girshick, R., Donahue, J., Darrell, T., Malik, J.: Region-based convolutional net- works for accurate object detection and segmentation. IEEE Transactions on Pat- tern Analysis and Machine Intelligence 38(1), 142–158 (2016) 3. Han, J., Zhang, D., Cheng, G., Liu, N., Xu, D.: Advanced deep-learning techniques for salient and category-specific object detection: A survey. IEEE Signal Processing Magazine 35(1), 84–100 (2018) 4. Ren, S., He, K., Girshick, R., Sun, J.: Faster R-CNN: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence 39(6), 1137–1149 (2016) 5. Ronneberger, O., Fischer, P., Brox, T.: U-Net: Convolutional networks for biomedi- cal image segmentation. In: Proceedings of the International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 234–241 (2015) 6. Lopes-dos Santos, R., Groot, R., Liying, S., Bossier, P., Van Stappen, G.: Halophilic bacteria as a food source for the brine shrimp Artemia. Aquaculture 500, 631–639 (2019) 7. Wang, G., Lopez-Molina, C., Vidal-Diez de Ulzurrun, G., De Baets, B.: Noise- robust line detection using normalized and adaptive second-order anisotropic Gaus- sian kernels. Signal Processing 160, 252–262 (2019) 8. Wang, G., Van Stappen, G., De Baets, B.: Automated detection and counting of Artemia using U-shaped fully convolutional networks and deep convolutional networks p. Under review (2019) 9. Wang, G., Van Stappen, G., De Baets, B.: Automated Artemia length measure- ment using U-shaped fully convolutional networks and second-order anisotropic Gaussian kernels p. Under review (2019) 10. Xing, F., Xie, Y., Su, H., Liu, F., Yang, L.: Deep learning in microscopy image analysis: A survey. IEEE Transactions on Neural Networks and Learning Systems 29(10), 4550–4568 (2018)