Effects of input-image size on performance of fish detection and species classification using deep learning

Deep learning has been extensively used in fisheries science, as it enables the acquisition of information regarding the body length and stock-abundance index of target fish from images, thereby facilitating stock assessment and management. However, generally, multiple species appear together in images obtained from fisheries, necessitating the classification of fish species before extracting relevant biological information. Improving the performance of fish detection and species classification is crucial as it affects the quality of biological information that could be inferred from images. Previous studies have reported that increasing the input-image size can affect the classification accuracy. Identification characteristics of fish are small in comparison with their body size, and increasing the image size can affect the classification accuracy; however, there are no reports on the effect of image size on fish species-classification accuracy. Herein, different input-image sizes were taken to evaluate the effect of input-image size on the performance of fish detection and species classification. Fish images (41,922 fish across 41 classes) were acquired from conveyor belts to sort set-net fish catches. Fish were detected and classified using a mask region–based convolutional neural network. The effect of input-image size on performance was examined using nine datasets in three image sizes of 1333 × 888, 2000 × 1333, and 2666 × 1777 pixels, obtaining an average mAP50–95 value of 0.586, 0.612, and 0.609, respectively. Larger image sizes offered improved performance compared with that of the smallest, averaging 0.026 and 0.023 improvements in mAP50–95 at two larger image sizes. However, when comparing the degree of improvement between image sizes of 2000 × 1333 pixels and 2666 × 1777 pixels under fine-tuning conditions, the former size resulted in higher performance. Performance was observed to improve for species with low performance at the smallest image size; therefore, we can say that increasing the input-image size is a simple and effective way for improving detection and classification performance for these species. • Input image sizes were varied to evaluate the effect of input image size on the performance of fish detection and species classification. • Larger image sizes offered improved performance compared with that of the smallest, with average improvements of 0.026 and 0.023 in mAP50–95 at two larger image sizes. • Performance was observed to improve for species with low performance at the smallest image size.