A Deep Learning Based System for Animal Species Classification

• Main Authors: ZHENG, PEIYU, ZHANG, XINSEN, WANG, SIXIANG, CHEN, GUOYI, YE, WENXIN, WANG, KAI
• Format: Patent
• Language: English
• Published: 11.10.2018
• Subjects: CALCULATING; COMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS; COMPUTING; COUNTING; HANDLING RECORD CARRIERS; PHYSICS; PRESENTATION OF DATA; RECOGNITION OF DATA; RECORD CARRIERS

With the rapid development of hardware and software technology, the thought of deep learning was introduced into neural network. Thus convolutional neural network (CNN), as an emulation of biological vision system, has become an active research direction in the filed of artificial intelligence and computer vision by recent years. Based on convolutional neural network theory and deep learning technology, this disclosed embodiment implements a digital image classification method for animal species. The whole process includes image capture and image preprocessing, in which images are divided into training sets, validation sets and test sets. Subsequently, the network structure is constructed, followed by the optimization of the parameters and super-parameters in this network to achieve the highest recognition rate. Convolution neural networks acts as the primary functional block, and the model parameters are tuned by mini-batch iterations. The merit of this invention is automatic feature extraction with convolution neural networks and parameter tuning by the training procedure, which prevents human intervention. 32x32x1 32x32x32 32x32x32 16x16x32 16x16x32 16x16x32 128 8x8x32 *-.6 conv3x3, 32 conv3x3, 32 maxpool2x2 dense conv3x3, 32 conv3x3, 32 maxpool2x2 stride (1, 1) stride (1, 1) stride (2, 2) stride (1, 1) stride (1, 1) stride (2, 2) fully connected Figure 1 TRAINING OPTIMIZE SET PARAMETERS PREPROCESS LEARNING PREDICTED ONE -HOT - TRAINGM SET OPTIMIZE ENCODE HYPERPARAMETERS COLLECT LABELS INPUT IMAGES Figure 2