Anti-shielding and multi-scale dead fish identification system and method based on deep learning

Abstract

The invention discloses an anti-blocking and multi-scale dead fish recognition system and method based on deep learning, and aims to improve a Faster RCNN recognition model for the difficulties of complex dead fish recognition background, large scale change, blocking and the like so as to adapt to a dead fish recognition scene in a complex underwater environment. The method comprises the following steps: carrying out contrast-limited adaptive histogram equalization processing on a to-be-recognized image to enhance the local contrast of a target; designing a multi-scale feature enhancement module to solve the problem of dead fish recognition in a large-scale range, designing an anti-shielding module based on an attention mechanism, highlighting a target area, and eliminating background interference such as other noise; finally, utilizing the rotating rectangular frame to represent the dead fish target, so the recognition precision of the dead fish in the dense scene is greatly improved.

Description

technical field [0001] The present invention relates to the application technical field of image recognition, in particular to underwater dead fish recognition technology, in particular to an anti-occlusion and multi-scale dead fish recognition system and method based on deep learning. Background technique [0002] In the process of fish farming, due to the presence of bacteria and parasites in the breeding water, or the excessive density of the water body and the poor fluidity of the water body, resulting in insufficient oxygen concentration, it is inevitable that fish will die during the breeding process. After the fish dies, they sink to the bottom of the water first, and the internal organs ferment to produce gas and then float to the surface. During this process, the dead fish may come into contact with other live fish or be eaten by live fish, resulting in the spread of pathogens. In order to prevent this problem, it is urgent to design a dead fish identification metho...

AI Technical Summary

Problems solved by technology

The above methods are all directly applied to the existing models, without considering and solving the following four practical problems and factors in a targeted manner, resulting in low recognition accuracy: (1) Underwater images usually have the characteristics of low contrast and low brightness, which are different from underwater images. There are large differences in the images and cannot be transferred directly; (2) After the dead fish sink to the bottom, there will be intra-class occlusion between the dead fish or between the dead fish and the live fish. At the same time, the dead fish will also be disturbed by aquatic plants, etc. (3) Due to differences in the category and volume of dead fish, the scale of dead fish varies widely; (4) Faster RCNN and other methods use horizontal rectangular boxes to represent the target position, but due to The bodies of dead fish are generally distributed in different directions. The horizontal rectangular frame will contain a lot of useless background information, and in dense scenes, the rectangular frames will also overlap, which is easy to be excluded by the post-processing stage.

Images