Statistical Inference of Fish Populations
from Deep-Learning Data.

Description | The DeepEcomar project, supported by the Govern de les Illes Balears, seeks at estimating fish populations with the help of deep learning techniques.

An instance segmentation convolutional neural network (Mask R-CNN) has been trained to detect different fish species in sub-aquatic images. From the output of this network, the number of specimens N of a given species can be obtained. However, this value may not match the actual number of fish N* in the scene, due to: a) missed or mislabeled detections (false negatives); b) wrongly detected fish (false positives).

In order to assess the performance of a neural network, the ratio of true positives with respect to the total number of detections (Precision, P) and the ratio of true positives with respect to the total number of fish (Recall, R) are computed for a set of images S. Ideally, if this set of images is representative of the actual data on which the network is applied, these values can be used to estimate the number of false positives and false negatives in a scene. However, it must be taken into account that to compute P and R, the set S must be manually labeled, and it is possible that it contains labeling errors (i.e. fish unlabeled or incorrectly labeled).

Given the setup described in the previous paragraphs, the goal of this problem consists in giving an estimate of N* (the actual number of fish of a given species in a scene), knowing the output N of the network (number of fish of this species detected by the network) and the precision (P) and recall (R) ratios, computed for a set which may contain incorrect labels and that may not correctly represent the input image.

Mathematical background | machine-learning, statistics.

Coordinators | José Luis Lisani and Bartomeu Coll, University of Balearic Islands.