Convolutional Neural Network Based on HOG Feature for Bird Species Detection and Classification

Abstract

This work is concerned with the detection and classification of birds that have applications like monitoring extinct and migrated birds. Recent computer vision algorithms can precise this kind of task but still there are some dominant issues like low light, very little differences between subspecies of birds, etc are to be studied. As Convolution Neural Network is a stateof-the-art method with respect to the accuracy of various computer vision related work like object detection, image classification, and segmentation, so CNN based architecture has been proposed to do the experiment for this work. Besides we applied Gaussian and Gabor filters for noise reduction and texture analysis respectively. Histogram of Oriented Gradient (HOG) has been utilized for feature extraction as it is a widely accepted method and it can extract features from all portions of the image. LeNet and ResNet are two good architectures of CNN. In our work, we used the HOG extracted features as input to implement LeNet and ResNet. A standard dataset is used for the experiment and we found that LeNet based CNN gives better results than other methods like ResNet based CNN, SVM, AdaBoost, Random Forest, (we used for the experiment) and other existing state-of-the-art proposed work as well. The experimental results using LeNet based CNN gives 99.6% accuracy with 99.2% F-score , and 96.01% accuracy with 94.14% F-score in detection and classification of birds respectively.