An efficient real-time marine object detection algorithm based on convolutional attention

Due to the influence of light and suspended particles in water, underwater image has many problems, such as details blur, color distortion and scale diversification, which are always difficult for underwater object detection. To solve the difficulty of low real-time performance and low recognition accuracy caused by low-quality images to traditional object detection algorithms. This paper proposes an improved YOLOv7 method. Its main idea is to use Ghost module to replace the traditional convolution block, and reduce computational complexity of this model through the idea of deep separation convolution. At the same time, SimAM is introduced in the feature fusion network to enhance the feature expression ability of the network. Finally, Mosaic data enhancement and transformer module are used in input part and Head part to comprehensively improve object recognition performance of the model. The results show that compared with the original YOLOv7, the precision of this algorithm is increased by 4.2%, mAP increase by 1.7%, fps is increased by 5.2 frames, and the number of parameters is effectively reduced, which can effectively meet the requirements of marine treasures detection.