LAUSR

Abstract Aquatic organism recognition is a core technology for fishing industry automation and aquatic organism statistical research. However, owing to absorption and scattering effects, images of aquatic organisms generally present poor contrast and color distortions, weakening the discriminative representations and decreasing the recognition accuracy. In this study, an inner feature and kernel calibration module is proposed for improving the recognition accuracy by aggregating informative features. Specifically, a set of features in one convolutional layer is split into two portions, each of which is fed to different flows. One flow contributes to emphasizing significant features, whereas the other is responsible for calibrating convolutional kernels. Consequently, the proposed module can effectively encode prominent features, and obtain dynamic convolutional kernels. Moreover, in view of a lack of aquatic organism examples, we collect 22,806 images of aquatic organisms and form a database for aquatic organism recognition containing 20 classes of common aquatic organisms. Finally, comprehensive experiments validate that the proposed module improves the performance of convolutional neural networks in a variety of recognition tasks, without any additional overhead. Specifically, the proposed module improves the top-1 accuracy to 95.7%, 97.1%, and 78.9% for the aquatic organism database and two public databases, respectively. Thus, this study could be beneficial for aquatic organism monitoring and automatic fishing, and can provide training data for other aquatic organism recognition methods.