Two datasets were used in this study. The first was the Barra-Ruler-445 (BR445) dataset used in [10] [11], and publically available via [12] originated from the [9] study. The second dataset was the Barra-Area-600 (BA600) dataset and released to public domain on publication of this work via [13]. In both datasets, each harvested barramundi fish (Asian seabass, Lates calcarifer) was digitally photographed and its weight was measured and recorded against the image file name. All images had a millimeter-graded ruler placed next to the fish, see Figure 1 for examples. The weights ranged 0.2 kg - 1 kg in BR445, and 1 kg - 2.5 kg in BA600. The image scales (in millimeters per pixel) were determined manually by measuring the number of pixels between the end points of the 300 mm ruler present in each image. The BR445 image scales were checked by the automatic ruler-scaling (RS2) algorithm [11]. The BA600 images were taken from the same distance hence they had the same scale.

The fins of the fish can contribute significantly to the total fish image area, see typical examples in Figure 1. At the same time the fins’ contribution to the fish mass is negligible. Therefore, ideally, only the fish-body area should be used to estimate the fish mass. For example, using the fish area without considering the fin tail was found to be more accurate when predicting the mass of Jade perch Scortum barcoo [4]. Furthermore, the fins are highly flexible and are more likely to change shape during harvesting, or be damaged and/or erode during the production growth cycle.

Segmentation of 200 images (100 from each dataset) into fish-body and background was done manually using the GIMP open-source software program. The

resulting fish-body binary masks were individually scaled to have the same scale of 1 mm per pixel. In this study all custom computer programs were written in Python programming language, which was also used to calculate the fish-body pixel areas. The obtained fish areas and the corresponding measured mass values were fitted via Equation (4) and results displayed in Figure 2. The fit achieved highly accurate R 2 = 0.9819 , and M A R E = 5.1 % , which were comparable to the corresponding results obtained on other fish species [4] [5] [6] [7]. Figure 2 clearly illustrated how the weight of the harvested Asian seabass Lates calcarifer could be estimated from the fish area with high accuracy. However, before such estimation method could be deployed in the aquaculture production environment, a robust automatic body-area extraction algorithm would be required, which was the focus for the rest of this section.

The recently developed semantic-segmentation Convolutional Neural Networks (CNN) [14] were highly successful in solving challenges where the segmentation of an image into per-pixel classes was required [11] [14] [15]. As discussed in the introduction, the second primary goal of this study was to design a practical Computer Vision algorithm to extract fish-body area from images. The Deep Learning neural networks [16] have revolutionized modern Machine Learning including the field of Computer Vision, and a large number of segmentation Deep Learning CNN models have been proposed. Comparing even the most popular segmentation CNN models was outside the scope of this work. Instead, the most accurate Fully Convolutional Network from [14], FCN-8s, was used. FCN-8s could be viewed as the modern baseline segmentation CNN model due to its highest citation rate out of all available segmentation CNNs (more than 4000 Google Scholar citations at the time of writing).

The FCN-8s model was implemented [17] in Python utilizing the high-level neural networks Application Programming Interface (API) Keras [18] together

with the machine-learning Python package TensorFlow [19]. The FCN-8s model is a general features-to-segmentation decoder CNN, which required an image-to-features CNN encoder. The original FCN-8s [14] was built with the VGG16 [20] convolutional layers as the encoder. The VGG16 model within Keras was trained to recognize 1000 different ImageNet [21] object classes and commonly referred to as ImageNet-trained. The ImageNet-trained CNN models were often more accurate than randomly initialized CNN models when they were further trained to recognize new object classes [22]. Therefore the convolutional layers of the ImageNet-trained VGG16 model were used to build our version of the FCN-8s model referred at the Fish Area Segmentation (FAS) model hereafter.

The FAS model was loaded with the relevant VGG16 weights facilitating the knowledge transfer [22], where the remaining convolutional as well as de-convolutional FCN-8s layers were initialized by the uniform distribution as per [23]. Furthermore, the first two FCN-8s decoder layers had their number of neurons reduced to 512 comparing to the 4096 neurons of the original FCN-8s in [14]. Such drastic reduction was justified by the requirement to recognize and segment only the single class of objects, i.e. fish body. The sigmoid activation function was used in the last layer.

The described 200 images together with the corresponding hand-segmented body masks were used to train the FAS. The 200 image-mask pairs were randomly split 80% - 20%, where the 80% of pairs were used as the actual training set and the remaining 20% were used as the validation set to assess the training process. Since the training set had such small number of images, the encoding VGG16 layers in FAS were fixed and excluded from training. The remaining trainable weights (excluding biases) were regularized by a weight decay set to 1 × 10 − 4 . The training and validation images as well as the masks were rescaled to 1mm per pixel. Then each image-mask pair was extensively augmented for each epoch of training, i.e. one pass through all available training and validation images. Specifically, the python-opencv package was used to perform augmentations, where each image and if applicable the corresponding binary mask were:

・ randomly rotated in the range of [−180, +180] degrees;

・ randomly scaled vertically in the range of [0.8, 1] and independently horizontally within the same range;

・ randomly cropped to retain 480 × 480 pixels;

・ each color channel was ±12.5 range randomly shifted;

・ randomly flipped horizontally and vertically;

・ ImageNet color mean values were subtracted as required when working with the VGG16 model.

To assist better segmentation, the following loss function was adopted,

l o s s ( Y g t , Y p r e d ) = 1 − d i c e ( Y g t , Y p r e d ) + b c ( Y g t , Y p r e d ) , (5)

where: Y p r e d and Y g t were the predicted and ground truth (i.e. segmented-by-hand) 480 × 480 masks; b c ( Y g t , Y p r e d ) was the standard binary cross-entropy; and where d i c e ( Y g t , Y p r e d ) was the Dice coefficient [24] ranged between zero and 1 (for identical Y p r e d and Y g t ). Since the sigmoid function was used as the last activation, the per-pixel predictions Y p r e d ranged between 0 and 1. The ground-truth Y g t was per-pixel encoded as zeros for the background pixels and ones for the body pixels. The training and validation losses were averaged over all pixels and all corresponding images obtaining the total training and validation losses for each epoch.

Keras implementation of Adam [25] was used as the training optimizer. The Adam learning-rate (lr) was set to l r = 0.001 , where the rate was halved every time the total epoch validation loss did not decrease after 16 epochs. The training was done in batches of 8 images, and was aborted if the validation loss did not decrease after 32 epochs, where the validation loss was calculated from the validation set of images and masks, which were not used by the optimizer for training the FAS model. While training, the FAS model with smallest running validation loss was continuously saved. Furthermore, if the training was aborted, it was restarted (from the previously saved FAS model) two more times with the initial learning rates l r = 0.5 × 10 − 3 and l r = 0.25 × 10 − 3 , respectively. Note that both the validation images were also augmented by the preceding augmentation pre-processing steps in order to prevent the indirect fitting of the validation images.