ASABE Technical Library - Abstract

Abstract.

The performance of machine learning methods is heavily dependent on the choice of features. Elaborate feature design is a way to take advantage of human ingenuity and prior knowledge. Such work is important but labor-intensive and highlights the weakness of many traditional learning algorithms. Recent findings have made the learning of deep layered hierarchical representations of data possible. This methodology has been devoted to learning algorithms for deep architectures such as Deep Belief Networks and Sparse Autoencoder based Neural Networks, with impressive results obtained in several areas, mostly on vision and language data sets. Based on the previous works and achievements, we established a dataset for the visual localization and classification of soft shell shrimp which have great similarity with sound shrimp. The identification work was completed by computing a discriminate map and application of deep learning based on Sparse Autoencoder based Neural Networks (SAENN). We construct a large standard shrimp dataset with 200 sound shrimp and 200 soft shell shrimp, a totally of 400 shrimp samples are contained in this dataset. The proportion of training samples and test samples approaches to 1:1. This shrimp classification training model was constructed based on self-learning of local image features. On the test stage of soft shell shrimp identification, our applications achieved a mean accuracy more than 98.05%, which indicates the deep learning algorithm is feasible to be used in identifying the soft shell shrimp on-line. Furthermore, some significant improvements will be implemented in the near future.