Research On Non-destructive Monitoring Of Egg Development Based On Semantic Segmentation And Illumination Method

In recent years,the per capita demand for poultry meat in China has steadily increased.Currently,the poultry hatchery industry has progressively adopted automated technology and equipment.The resourceintensive and labor-intensive hatching industry is developing towards fine,intelligent,and information-based hatching.Nevertheless,the majority of egg development detection is still performed manually.This method requires employees to inspect hatching eggs with an egg testing lamp and observe their development with the naked eye.Even seasoned employees are susceptible to visual fatigue,which can result in ignored or incorrect inspections.When unfertilized and dead eggs are mixed with viable eggs,it not only wastes resources but also leads to spoilage and bacterial growth,which has a negative impact on the hatching rate.Moreover,air-cells change during the development of eggs,and accurate monitoring of the hatching egg incubation process can be achieved through accurate determination of air-cell changes in hatching eggs.This thesis focuses on the monitoring of air cell changes and uses LED lighting to collect egg-candling images with the obvious areas of air-cells based on the different internal light transmittance properties of the eggs.An air-cell dataset of egg-candling images was created by manually annotating and extracting egg-candling images collected from days 7 to 19.With the improved DeeplabV3+ semantic segmentation network,training and adjusting parameter were performed to achieve the goal of the air-cell segmentation accurately.Based on DeeplabV3+ semantic segmentation network and combined with the application scenario of industrial production line,the lightweight network Mobilenet V2 with optimized structure is used instead of the original backbone network to improve the operating efficiency.In view of the characteristics that the detailed information of air-cells is more and the position of the air-cells are relatively fixed,the attention mechanism is embedded in the backbone network to enhance the network feature expression capability and improve the feature spatial relationship for the features.To further make full use of the semantic information of the input egg-candling image,a new decoder is proposed.It uses semantic embedding branching to fuse the high,medium and low-level features extracted from the encoder,which makes the model more robust and interpretable.To improve the problem of positive and negative sample imbalance caused by small aircell targets,the Focal Loss optimized loss function is introduced during training.For the problem of the insufficient data of egg-candling images,transfer learning is used by employing the pre-trained weights of the backbone network on the Image Net dataset to further improve the accuracy.The experimental results demonstrate that the proposed enhanced design strategy can substantially improve the segmentation accuracy and running efficiency of air-cells in egg-candle images.When the lightweight Mobilenet V2 with an embedded convolutional block attention module is used for the backbone network,could reduce model parameter size from 208.72 MB to 26.50 MB,increase running speed from24.3 FPS to 47.7 FPS,and the algorithm has a m Io U of 87.24 % and a m PA of 93.36%.It maintains a high level of segmentation precision while enhancing algorithm performance.On this basis,it employs the new decoder and the coordinate attention with less computational complexity,as well as the transfer learning method.Then,the m Io U increases to 89.06%,the m PA increases to 94.66%,and the running speed is 40.4FPS.This method achieves the high segmentation accuracy and the perfect operation efficiency,and the hatching rate of the eggs in this study is normal,which satisfies the requirements of non-destructive testing.Finally,the feasibility of the proposed method is verified by monitoring from 7 to 19 days through application during the development of egg.