Manano, M. | Search bioRxiv

bioRxiv2024

Effects of input image size on the accuracy of fish identification using deep learning.

The length composition of catches by species is important for stock assessment. However, length measurement is performed manually, jeopardizing the future of continuous measurement because of likely labor shortages. We focused on applying deep learning to estimate length composition by species from images of fish caught for sustainable management. In this study, input image sizes were varied to evaluate the effect of input image size on detection and classification accuracy, as a method for improving the accuracy. The images (43,226 fish of 85 classes) were captured on conveyor belts to sort set-net catches. Fish detection and classification were performed using Mask R-CNN. The effect of input image size on accuracy was examined using three image sizes of 1333x888, 2000x1333, and 2666x1777 pixels, achieving an mAP50-95 of 0.580 or higher. The accuracy improved with increasing image size, attaining a maximum improvement of 4.3% compared to the smallest size. However, increasing the image size too far from the default size may not improve the accuracy of models with fine-tuning. Improvements in accuracy were primarily observed for the species with low accuracy at the smallest image size. Increasing image size would be a useful and simple way to improve accuracy for these species.

ecology↗

bioRxiv2023

Length estimation of fish detected as non-occluded using a smartphone application and deep learning techniques

Uncertainty in stock assessment can be reduced if accurate and precise length composition of catch is available. Length data are usually manually collected, although this method is costly and time-consuming. Recently, some studies have estimated fish species and length from images using deep learning by installing camera systems in fishing vessels or a fish auction center. Once the deep learning model is properly trained, it does not require expensive and time-consuming manual labor. However, several previous studies have focused on monitoring fishing practices using an electronic monitoring system (EMS); therefore, it is necessary to solve many challenges, such as counting the total number of fish in the catch. In this study, we proposed a new deep learning-based method to estimate fish length using images. Species identification was not performed by the model, and images were taken manually by the measurers; however, length composition was obtained only for non-occluded fish detected by the model. A smartphone application was developed to calculate scale information (cm/pixel) from a known size fish box in fish images, and the Mask R-CNN (Region-based convolutional neural networks) model was trained using 76,161 fish to predict non-occluded fish. Two experiments were conducted to confirm whether the proposed method resulted in errors in the length composition. First, we manually measured the total length (TL) for each of the five fish categories and estimated the TL using deep learning and calculated the bias. Second, multiple fish in a fish box were photographed simultaneously, and the difference between the mean TL estimated from the non-occluded fish and the true TL from all fish was calculated. The results indicated that the biases of all five species categories were within {+/-} 3%. Moreover, the difference was within {+/-} 1.5% regardless of the number of fish in the fish box. In the proposed method, deep learning was used not to replace the measurer but to increase their measurement efficiency. The proposed method is expected to increase opportunities for the application of deep learning-based fish length estimation in areas of research that are different from the scope of conventional EMS.

ecology↗