Effective Training of Convolutional Neural Networks for Insect Image Recognition

摘要

Insects are living beings whose utility is critical in life sciences. They enable biologists obtaining knowledge on natural landscapes (for example on their health). Nevertheless, insect identification is time-consuming and requires experienced workforce. To ease this task, we propose to turn it into an image-based pattern recognition problem by recognizing the insect from a photo. In this paper state-of-art deep convolutional architectures are used to tackle this problem. However, a limitation to the use of deep CNNs is the lack of data and the discrepancies in classes cardinality. To deal with such limitations, transfer learning is used to apply knowledge learnt from ImageNet-1000 recognition task to insect image recognition task. A question arises from transfer-learning: is it relevant to retrain the entire network or is it better not to modify some layers weights? The hypothesis behind this question is that there must be part of the network which contains generic (problem-independent) knowledge and the other one contains problem-specific knowledge. Tests have been conducted on two different insect image datasets. VGG-16 models were adapted to be more easily learnt. VGG-16 models were trained (a) from scratch (b) from ImageNet-1000. An advanced study was led on one of the datasets in which the influences on performance of two parameters were investigated: (1) The amount of learning data (2) The number of layers to be finetuned. It was determined VGG-16 last block is enough to be relearnt. We have made the code of our experiment as well as the script for generating an annotated insect dataset from ImageNet publicly available.