In order to achieve efficient identification and classification of tuna longline catches and to improve the accuracy of fishery resources monitoring, this study explores a fish image recognition method based on convolutional neural network. The experiments were conducted using image data of three economic fish species and ten bycatch species caught by the Songhang of Shanghai Ocean University during the high seas survey in the western and central Pacific Ocean, and a convolutional neural network （CNN） based on a single shot multiBox detector （SSD） was used to classify and recognise the images. The training dataset is optimised by comparing and analysing the local fish images with the overall image dataset to improve the classification performance of the model. The experimental results show that the classification accuracy of the improved fish image dataset on the SSD model reaches 91.6%, which is a 6.2% improvement compared to the original dataset. The study shows that using the optimised dataset, the SSD model can significantly improve the recognition accuracy of tuna longline catches with better stability and adaptability. This study provides an effective technical path for fishery resource monitoring based on convolutional neural networks, especially in improving the automatic classification and identification accuracy of tuna longline catches, which is of great significance for promoting sustainable fishery management and marine ecological protection.