SPECIES DETECTION USING DEEP LEARNING FOR BIRDS

Abstract

Some bird species are becoming more uncommon, and even if they are discovered, it may be difficult to classify them. From a human viewpoint, birds may be seen in a variety of sizes, shapes, colours, and angles. Because of this, it is easier to identify birds by looking at their photos, as opposed to listening to their calls. In addition, it is easier for humans to identify birds based on their appearances in photos. As a result, our technique relies on the Caltech-UCSD Birds 200 dataset for both training and testing. To produce an autograph using tensor flow and a deep convolutional neural network (DCNN), a picture is transformed to grey scale using the DCNN technique. The testing dataset is used to compare the various nodes, and a score sheet is generated as a result. The highest score on the scoring sheet may be used to determine which bird species is needed. Analyses of the dataset (Caltech-UCSD Birds 200 [CUB-200-2011]) have shown that the system is 80-90% accurate in identifying birds. The experiment is run on Ubuntu 16.04 using a Tensor flow library on a laptop.