Effect of Image Resolution on CNN Classification Performance: An Experimental Study on Sunflower Seed Classification

Deep learning-based image classification systems, especially CNN architectures, have achieved significant successes in recent years. However, it is not only the architectural design and training strategies that affect model performance; the optical hardware used in the data collection process is also a determining factor. This study aims to systematically examine the impact of physical image resolution on CNN performance. As a case study, the sunflower seed classification problem was chosen, where classification between classes is difficult. Images of the same seeds were captured at two different resolutions using 6 mm and 12 mm lenses. MobileNetV3L and ResNet152 models were used for train and test the datasets. With the ResNet152 model test accuracy achieved as $93.16 \%$ with 6 mm lens dataset and $97.72 \%$ with 12 mm dataset. The highest test accuracy was calculated as $99.14 \%$ with the MobileNetV3L model using the highresolution 12 mm dataset. The highest increase of $4.56 \%$ was observed in the ResNet152 model (from 93.16% to 97.72%). Results shown that high-resolution images significantly enhance classification performance. The findings indicate that optimizing imaging hardware is as important as the model architecture. In this context, the obtained results indicate that artificial intelligence systems should be optimized not only at the algorithm level but also in terms of imaging hardware.