Research And Implementation Of Automated Fish Fry Counting Method Based On Deep Learning

Fry fish counting is an important problem for the aquaculture industry.Traditional methods that count fry fish through weighting them are often with a very low level of accuracy.Although a number of image threshold segmentation methods can be used for counting fry fish automatically,they are unsuitable for high-density environments where targets are too close together.Our aim in this paper is to address these limitations by improving the accuracy and speed of fish fry detection in complex environments by the utilization of deep learning techniques.In particular,we propose a fish fry counting method based on deep learning target detection technology that can effectively mark fish fry in an image and allow users to count it.Our proposed algorithms are optimized for small-target detection in high-density backgrounds and can improve the accuracy and speed of fish fry detection in real-world applications.The following are our main contributions:1.We developed a Single-type Small Target Tracking algorithm,called the SST algorithm,to improve the efficiency of small target detection.As part of the SST algorithm,we developed a shallow neural network that can output more small target information feature maps.In addition,we designed a loss function that can accelerate model training.2.Aiming at the problem of fish fry detection redundancy in a high-density environment,we propose an Iterating SST(Iter-SST)algorithm that can be used in a highdensity environment on the basis of the SST algorithm.The Iter-SST algorithm uses an iterative method to achieve target detection from easy to difficult.In parallel with this,a new non-maximum suppression algorithm soft-DIo U-NMS(Soft Distance-Io U Non-Maximum Suppression,soft-DIo U-NMS)is proposed,which takes the distance between the center point of the predicted frame and the real frame as a threshold,recalculates the scores of the detected frames,and gradually removes the frames with lower scores.In this paper,a smart fish fry counting device is developed based on the Iter-SST algorithm,which takes pictures of fish fry and uploads them to a server,where the pictures are processed and counting results are returned.The device has been tested repeatedly at the Shatang snakehead fry breeding base of Anhui Junlong Wangstate Agriculture Co.,Ltd.The test results show that this smart fry counting device can control error within 4%,which is a significant improvement in counting accuracy.