Improving Dog Image Classification with CNN Model and Image Distance method

This research focuses on the recognition and retrieval of various dog breeds based on their distinct characteristics. The retrieval process involves identifying and extracting relevant images from a comprehensive dataset in response to a query image. This method leverages the analysis of image features such as shape, color, and texture to accurately classify images into different dog breeds. Dog imaging serves as a specialized application that is beneficial for breed identification and holds potential in scenarios like lost dog recovery and image-based adoption systems. Despite significant advancements in general image search methodologies, the unique challenges associated with dog image retrieval remain underexplored. In this study, several Convolutional Neural Network (CNN) models are employed for feature extraction from dog images. The models utilized include InceptionResNetV2, XCeption, DenseNet121, InceptionV3, NasNetMobile, ResNet50V2, and MobileNetV2. These CNN models are evaluated using different similarity measures— Euclidean, Manhattan, Cosine, and Minkowski—to enhance retrieval accuracy. The proposed approach is tested on a large dataset encompassing diverse breeds and environmental conditions. Experimental results demonstrate the effectiveness of various CNN architectures in extracting pertinent features and highlight how each similarity measure influences search performance.The findings reveal that the InceptionResNet model, combined with Euclidean and Manhattan similarity measures, achieves an average accuracy of 90%, followed by the NasNetMobile and Xception models with an average accuracy of 86%. These results underscore the strengths and limitations of each model and provide valuable insights for future research and practical applications in canine image retrieval systems.