Quantum-enhanced convolutional networks for animal image classification in environmental monitoring: a comparison with Vision Transformer and VGG16

Abstract

Automated recognition of animals in images supports wildlife surveys, livestock monitoring and other environmental applications, and it needs classifiers that stay accurate on varied images. This study compares three deep learning models for image classification on the CIFAR-10 benchmark: a quantum-enhanced convolutional neural network (EQ-CNN), a Vision Transformer (ViT) and VGG16. CIFAR-10 contains six animal classes (bird, cat, deer, dog, frog, horse) and four vehicle classes, so it allows animal and non-animal classes to be compared under identical conditions. All images were resized to 224 × 224 pixels and normalised, and each model was trained for five epochs. The EQ-CNN combines convolutional layers with quantum-inspired non-linear activations. ViT encodes images with self-attention, and VGG16 is the classical convolutional baseline. Performance was assessed with accuracy, precision, recall, F1-score, confusion matrices, ROC curves, RMSE and R². K-means clustering and class similarity graphs were used to examine feature separation. The EQ-CNN scored highest in every class; its mean per-class score over the six animal classes was 0.945, against 0.913 for ViT and 0.877 for VGG16. The results suggest that quantum-inspired hybrid designs are worth testing on real animal-monitoring data such as camera-trap images. Performance was evaluated in terms of accuracy, precision, recall, F1-score, confusion matrices, ROC curves, RMSE and R². The K-means clustering and class similarity graphs were used to analyze feature separation. The EQ-CNN achieved the highest mean per-class score (0.945) across all the animal classes, compared to ViT (0.913) and VGG16 (0.877). The results suggest that hybrid designs inspired by quantum should be tested on real animal monitoring data, for example on camera-trap images.