emoen/Deep-learning-for-salmon-scales

Fish scales constitute a valuable source of information about individual life histories, but correctly extracting this information requires a highly skilled expert. Here, we train a deep convolutional neural network architecture EfficientNet B4 on a set of about 9000 salmon scale images, and show that it attains good performance on predicting a set of variables used in stock management. Further, we see substantial benefits from user transfer learning with a network pre-trained on ImageNet, even if the salmon scale images are very different from those found in the data used for pre-training.