Fish quantity detection method based on deep neural network

Abstract

The invention provides a fish quantity detection method based on a deep neural network, and belongs to the technical field of underwater computer vision, and the method comprises the steps: obtaining a target net cage fish school image, carrying out the feature enhancement processing, taking the image as an image data set, marking fishes in the image data set, obtaining a data set of a real value density map, and obtaining a real value density map; and constructing a fish school density estimation model based on a multi-channel attention mechanism, re-establishing an image data set of each depth layer of the target net cage, and respectively inputting the image data set into the fish school density estimation model to obtain a fish school quantity statistical result and a density map of each depth layer of the target net cage. According to the method, the target features are clearer and more prominent by performing feature enhancement on the fish school image, meanwhile, the suspended particle features in the background are weakened, so that the interference of suspended particles on fish school recognition is reduced, a multi-channel attention mechanism is introduced into a fish school density estimation model, the recognition capability of the model on overlapped fish schools is enhanced, and the recognition accuracy is improved. And the quantity statistics accuracy is improved.

Images