DQFCNN: A New Deep Learning Architecture for Bird Species Classification Based on Images

Bird classification and identification are very important for protecting birds. Birds come in a wide range of sizes, shapes and colors and their appearance can even change depending on the angle. Indeed, the images exhibit various distinctions that should be documented for audio recognition of bird species. Additionally, it is easier for people to identify birds in the pictures. However, identifying bird species is difficult due to the similarities existing between the intra-class and inter-class varieties of bird species. To overcome the issues, a Deep Learning (DL) Architecture named Deep Quantum Fusion Convolution Neural Network (DQFCNN) is implemented. The input images from the dataset are pre-processed using the Kalman filter, which removes unnecessary noise. Then, bird segmentation is accomplished by Fast Fuzzy C-Means (FCM) clustering, which is followed by image augmentation that includes resizing, rotation, translation and random erasing. Afterward, features, such as Local Vector Pattern (LVP) as well as statistical features are extracted from the augmented image. Here, the use of the Kalman filter, bird segmentation and comprehensive image augmentation techniques further contributes to the robustness of the model. Finally, the proposed DQFCNN is processed and integrates the Convolutional Neural Network (CNN) model and Deep Quantum Neural Network (DQNN) to classify Indian bird species. This fusion leverages the strengths of both classical and quantum computing paradigms to enhance classification accuracy. In this work, the Indian-Birds-Species-Image-Classification dataset was used for both training and validation in this technique. In addition, the overall performance of the proposed DQFCNN method for Bird Species Classification is compared with other existing methods, such as Perceptual, Descriptive and Harmonic Features (PDHFs), Batch Normalization Dropout Fully-Connected (BNDFC), Deep Neural Network (DNN), Convolutional Neural Network+Recurrent Convolutional Neural Networks (CNN+RNN), Wav2vec, Deep Convolutional Neural Networks (DCNNs) Fast Shape-based Network (FS-NET) and Phylogenetic Perspective Neural Network (PPNN). The performance of the devised DQFCNN is analyzed by performance metrics, including an accuracy of 87.6%, True Positive Rate (TPR) of 94.7% and the smallest False Positive Rate (FPR) of 81.6%. Additionally, the proposed model demonstrates superior accuracy, outperforming other methods by 11.45% over PDHF, 8.90% over BNDFC, 6.47% over DNN, 4.39% over CNN+RNN, 3.12% over Wav2vec, 2.31% over DCNN, 1.85% over FS-NET and 1.27% over PPNN.