Advanced Marine Object Identification Approach Utilizing an Enhanced YOLOv7-Tiny Detection Model

The global aquaculture industry is booming because more and more people want sea treasures like sea cucumbers, sea urchins, and shells. But the fact that underwater environments are so complicated makes it hard for people to find and identify sea treasures during the traditional manual inspection process. To solve this problem, this paper suggests a target detection algorithm for underwater sea treasures called UPA-YOLO. Its goal is to accurately and quickly find underwater treasures and speed up the inference process through model transformation so that the detection model can be used on edge devices. First, the MAFPN neck structure is used to replace the ELAN structure based on the YOLOv7-Tiny network. This allows for the multi-scale capture of semantic information about underwater sea treasures and improves the UPA-YOLO model's ability to accurately find these treasures. Second, the P2ELAN module is built and added to the backbone network. It uses the redundancy information in the feature map and dynamically adjusts the convolution kernel to fit the data. The P2ELAN module is added to the backbone network. It uses the extra information in the feature map to dynamically adjust the convolutional kernel to make up for the lack of data, which reduces the number of parameters in the model. It also adds the MSCA attention mechanism to block the complex and changeable background features underwater, which improves the UPA-YOLO model's ability to extract semantic features from underwater targets. The MPDiou loss function is added to the improved algorithm model, and the data validation of the detection model is finished. Finally, the target detection is optimised based on the TensorRT acceleration framework. Finally, the target detection model is optimised using the TensorRT acceleration framework. The Jetson Nano edge device is then used to finish the localisation deployment and make the real-time target detection task of underwater sea treasures possible. Our optimised algorithm achieves a mAP0.5 of 79.5%, which is a 2.0 percentage point improvement over the baseline YOLOv7-tiny model (77.5%). It also keeps the computational efficiency at 45.2 GFLOPs. The model gets an inference speed of 86 ms per image when it is used on Jetson Nano edge devices. This is four times faster than the unoptimised version. This level of performance makes it possible to find marine products in the field in real time.