<p>Manual determination of fish age using hard tissues demands considerable effort. However, reliable and high-performance automatic alternative methods are not widely available. The aim of this study was to investigate the influence of input image size on the performance and the image areas used for automatic age determination. To this end, we used input images with sizes of 240 × 240–960 × 960 or 170 × 340–679 × 1358 pixels, either whole or trimmed in half, of 3- to 5-year-old chum salmon scales (<i>n</i> = 1835), as well as deep convolutional neural networks (CNN). In entire-scale images, high accuracy was achieved when the input image size exceeded 679 × 679 pixels, reaching a maximum of 94.7%. Below this size, accuracy decreased significantly, and overfitting became pronounced. At sizes exceeding 480 × 480 pixels, the CNN consistently based its determinations on areas outside the first annulus, similar to visual inspection. Conversely, at sizes below 480 × 480 pixels, where the circuli become indistinguishable, the CNN focused on a wide range around the focus. In trimmed-scale images, CNN accuracy plateaued at a lower level (80.8–88.7%) than in entire-scale images, even for large-sized images. These results suggest that using sufficiently large entire-scale images which retain detailed information about the circulus pattern is important for achieving high performance of frameworks for automatic age determination.</p>