Study On Detection System Of Population Cotton Seed Skin Damage Based On Deep Learning

Xinjiang is the most important cotton producing area in China.Screening high-quality cottonseed is the key to improve cotton yield.Detecting the damage of cottonseed skin is one of the important indexes to evaluate the quality of cottonseed.High-quality cottonseed can improve the yield and quality of cotton,which is of great significance for the development of cotton industry and agricultural production.Therefore,in order to screen out high-quality cottonseeds,this paper takes Xinjiang ’s ’ Xinluzao-78 ’ cottonseeds as the research object,combined with the non-destructive testing technology widely used in the current algorithm model,and proposes a fast and accurate method for detecting the epidermal damage of group cottonseeds to meet the actual production needs.The main work of this paper is as follows :(1)By analyzing the needs of the group cottonseed damage detection process,a group cottonseed image acquisition platform was designed and built,and the group cottonseed image information of’Xinjiang Xinluzao-78 ’ was collected.A two-category data set of 500 non-destructive cottonseeds and damaged cottonseeds was constructed.At the same time,the collected group cottonseed image data was expanded to 1500 by data enhancement,and the data was labeled according to the feature information of damaged and non-destructive cottonseeds.Finally,it was assigned to a training set and a test set for the detection of group cottonseed skin damage.(2)Optimization and improvement based on SSD-VGG16 algorithm model.Firstly,the SSD backbone network VGG16 is replaced by Mobilenet-v2.Secondly,four(Inverted Residual＿1 ～ 4)reverse residual modules and a convolution layer are added after the backbone network to deepen the semantic information.Finally,the group cottonseed data set is trained by transfer learning.The results show that the improved SSD-Mobilenetv2 algorithm model has a 77.47 % accuracy rate for group cottonseed damage.Compared with SSD-VGG16 algorithm model,although there is a slight decrease in detection accuracy,the memory consumption and detection speed are better than SSD-VGG16.The improved algorithm model has a detection speed of 0.051 s for a single image.(3)Optimization and improvement based on YOLOv5 s algorithm model.Firstly,the number of modules in the YOLOv5 s backbone network layer is simplified,and the redundant information is reduced.The Denseblock module is used to replace Focus to improve the feature extraction ability of the network.Secondly,after the SPP pooling layer,the CA collaborative attention mechanism module is added to reduce the large target detection layer and guide the network to pay more attention to the location,channel and dimension information of small targets.Then,Ghost Conv is used in Neck instead of the traditional convolution layer to reduce the amount of floating-point calculation and speed up the reasoning speed of the model.Finally,the CIOU loss function is used as the boundary regression loss function to improve the recall rate of the model.The improved network model is used to train the group cottonseed data set.Combined with the transfer learning method,the accuracy rate,recall rate,m AP,priori box loss,classification loss and target loss are used as the evaluation indexes of group cottonseed detection.Through ablation experiment verification analysis and comparison with YOLOv4,YOLOv5 s,SSD-VGG16 network model,the experimental results show that the improved YOLOv5 s network model can still effectively identify group damaged cottonseeds while greatly reducing the amount of floating point calculation,and the m AP value is 6.4 % higher than that of the original YOLOv5 s.The model recognition accuracy,recall rate and m AP were 92.4 %,91.7 % and 98.1 %,respectively.(4)The improved YOLOv5 s algorithm model is deployed on the Windows side using the Py Qt5 platform development tool to implement the UI programmatic interface.The single frame image detection speed is only 0.028 s with the support of NVIDIA Ge Force RTX 3060 graphics card.The stability and reliability of the algorithm are verified by the real-time detection of the tracked test platform.Compared with SSD and YOLOv5 s network models,the improved YOLOv5 s network model solves the problem of missed detection,reduces false detection and has less floating-point calculation,and the network memory occupation is only 9.88 MB.It can run smoothly on the Windows side and meet the needs of real-time detection,and the improved model has good generalization and robustness.