Multi-Level CNN Feature Fusion from ResNet50 for Near-Duplicate Image Detection in Real Estate Imagery

Abstract

The volume of images uploaded to the internet is increasing at an unprecedented rate, making image deduplication, through accurate near-duplicate detection, a critical task in computer vision. However, comparing images for similarity remains challenging due to complex visual structures and subtle appearance variations.We propose a novel embedding method for image similarity detection. It constructs an enriched representation by concatenating outputs from multiple intermediate layers of a pre-trained ResNet50 convolutional neural network and trains a lightweight decision network on top to classify image pairs. Unlike aggregation approaches that average or sum intermediate features, our method preserves both low-level and high-level information in a single descriptor and maintains feature diversity. The multi-level embedding is further normalized to balance feature contributions and is evaluated against classical keypoint descriptors, a DCT-based perceptual hash, and a standard single-layer ResNet50 embedding.We evaluate this method on three real-world image deduplication tasks derived from real estate listings, covering (a) near-identical property photos with graphical overlays, (b) interior room photographs captured from different angles, and (c) schematic floor plan images. The proposed embedding achieves F1-scores of 0.96, 0.87, and 0.77, representing a 10-15% absolute improvement over baseline methods, including classical feature descriptors and standard ResNet50 final-layer embeddings.This approach has been successfully deployed in production on a large-scale real estate platform, reducing duplicate images and improving search quality. The results demonstrate that multi-layer CNN embeddings with explicit feature preservation offer a robust and scalable solution for near-duplicate image detection in structured domains such as real estate photography and schematic floor plans.