We present an example of a technical tool for transforming a huge volume of noisy information into useful ecological data. Modeling fish population dynamics is challenging and accurate predictions are essential for providing proper scientific assessment on fishing quotas and fishery management. Here we present the preliminary results of a research project aimed at automatically measuring fish lengths, which are a key input for feeding conventional fishery models. A pre-trained deep convolutional network (Mask-RCNN) has been fed with more than 2000 patterns of fish heads, with the goal of detecting and segmenting all the instances of a particular species (hake) in the fish cages. Although only the heads of the fishes are detected by the network, in this communication we show how the total length of the fishes may be estimated based on the segmentation results, and we assess the statistical relevance of the obtained measurements.