EARLY PREDICTION OF 28-DAY COMPRESSIVE STRENGTH OF RECYCLED-AGGREGATE CONCRETE FROM FRESH-MIX SMARTPHONE IMAGES: COMPARATIVE EVALUATION OF TEN DEEP VISION ARCHITECTURES

Abstract

Early estimation of 28-day concrete compressive strength from the visual state of a fresh mixture could provide a rapid screening tool before conventional curing and destructive testing are completed. This study investigates an image-based prediction framework for concrete containing recycled aggregates derived from crushed pavement blocks that originally had a silica-rich wearing surface. Fifty-four concrete mixtures and 108 cube specimens were produced. Fresh concrete was photographed less than one minute after mixing with a Samsung Galaxy S23 Ultra fixed 30 cm above a tray under ring-light illumination. One hundred and fifty photographs were acquired for each mixture, yielding 8,100 raw images. The final benchmark comprised 108 selected representative images associated with the specimen records. Ten ImageNet-pretrained architectures—ConvNeXt-Small, ConvNeXt-Tiny, DenseNet121, EfficientNet-B0, MobileNetV3-Large, MobileNetV3-Small, ResNet18, ResNet34, ResNet50, and ViT-B/16—were evaluated using transfer learning, two-stage fine-tuning, and three random seeds (42, 123, and 2024). ConvNeXt-Small achieved the lowest mean validation RMSE (3.299 ± 0.039 MPa), followed by ConvNeXt-Tiny (3.448 ± 0.168 MPa). ViT-B/16 showed the highest seed sensitivity and the greatest computational demand. The findings support the feasibility of extracting strength-related information from standardized fresh-concrete smartphone images. The limited number of independent mixtures, the absence of a documented reproducible rule for selecting the final 108 representative images from the raw image pool, and the lack of external validation currently restrict generalization beyond the studied materials.