Bird Species Identification from Image Processing

Although viewing birds is a popular pastime, bird books are necessary to identify the species. We created a deep learning platform to help users identify 27 species of Taiwanese endemic birds using a smartphone app called the Internet of Birds (IoB), giving birdwatchers a convenient way to appreciate the beauty of birds. A convolutional neural network (CNN) was trained to identify key elements in bird photos. To balance the distribution of bird species, we first created and refined a circumscribed region of interest to fine- tune the object granularities' colors and shapes. To enhance feature extraction, the outputs of the prior and current layers were then linearly combined using a skip connection technique. Lastly, a probability distribution of bird traits was obtained by applying the softmax function. Images submitted by mobile users were identified using the learned criteria of bird traits. For the training images, the suggested CNN model with skip connections had a higher accuracy of 99.00% than the CNN's 93.98% and the SVM's 89.00%. The average sensitivity, specificity, and accuracy for the test dataset were 93.79%, 96.11%, and 95.37%, in that order. Key Words: Bird image recognition, convolutional neural network, deep learning, mobile app.