GRAPHS TARGET RELATIONS AND CLASS ATTENTION MODEL FOR DEEP BIRD SPECIES DETECTION NETWORK (Article)

Abstract

Birds are vital for ecosystem balance and scientific understanding due to their diverse traits and sensitivity to environmental changes. Fine-grained image analysis (FGIA) is crucial for bird classification, relying on discriminative feature representation and accurate localization. However, many existing models emphasize extracting category-specific features to reduce overlap in semantic features across classes, enhancing class differentiation. So, representing class-specific regions for large species is challenging task. In this paper, Class Attention based Deep Bird species detection network (CADepBnet) is developed to discriminate the sub-class of birds from the similar super-class for the rapid detection of various bird species effectively. In this method, Adaptive Graphs of Target Relations (AGTR) is constructed as an auxiliary feature regularization that is readily incorporated as Class Attention Layer (CAL) into Inception-ResNet-v2 to represent the class-specific regions of bird species. A non-parametric feature regularization method is introduced for deep representation learning for categorization that reveals latent associations across target classes. AGTR is constructed using an online center loss representations reducing variation within each class and detecting connections across the classes, thereby increasing the number of features for discrimination. Semantic interpolation based data augmentation provides a new approach to investigate AGTR. More turning images are generated by spatially transforming two arbitrary input images from distinct classes. A variety of target-image soft labels creates a more complex association graph. This AGTR is integrated with Inception-ResNet-v2 as CAL with addition class-wise feature map and spatial-wise pooling to derive the attention encoded feature representation for specific classes. The CAL is inserted with fully connected (FC) layer before the softmax layer. The assembled feature vector is fed into the FC that predicts the proper bird label for bird species identification using a feature map matrix and input weights. The proposed CADepBnet model attains 94.13% accuracy for CUB-200-2011 Dataset, outperforming Fine-Grained Bird Classification using Attention and Decoupled Knowledge Distillation (FGBC-ADKD), Progressive Cross-Union Network (PC-Net) and Attention Feature Enhancement Module - Siamese network (AFEM-SN). It also records 99.18% on Birds 525 Species Dataset compared to EfficientNetB0, EfficientNetB5 and ResNet-152 and achieves 96.65% on Indian-Birds-Species-Image-Classification Dataset, surpassing popular deep learning models.