Surface Defect Detection Method For Biscuits Based On Deep Learning

With the rapid development of the food processing industry,people’s demands for food safety and quality are increasing.Biscuits may have various surface defects during the production process,such as cracks,notches,and foreign objects,which can affect the taste and appearance of the biscuits and even pose a threat to consumer health.Biscuit surface defect detection is an important means to solve this problem,but there is relatively little research in this field.Therefore,this thesis proposes a deep learning-based method for detecting surface defects of biscuits,which can detect various surface defects of biscuits with high accuracy,stability and reliability,providing an effective quality inspection method for biscuit production.The main contents of this thesis are as follows.(1)This thesis establishes a dataset for surface defects of biscuits,collecting five categories of surface defects,including cracks and notches,as well as foreign objects such as minerals,hair,and metal.In order to improve the quality of the dataset,noise reduction processing is performed during image preprocessing,and data augmentation techniques are used to increase the quantity and diversity of the dataset,which facilitates the training and evaluation of deep learning models.(2)To verify the feasibility of deep learning models on the custom dataset,this thesis proposes methods for detecting biscuit surface defects based on Faster R-CNN,SSD,and YOLOv5 networks,comparing the detection accuracy,mean detection accuracy,model size,and speed of the three models through experiments and analyzing these methods.Experimental results show that the YOLOv5 algorithm performs best in the detection of surface defects of biscuits,thus serving as the basis for further research in this thesis.(3)In order to further improve the effectiveness of biscuit surface defect detection,this thesis uses the YOLOv5 model as the base network,and adopts the k-means++ clustering method to optimize anchor boxes,aiming to solve the problem of extreme aspect ratios and diverse pose variations of biscuit surface defects,in order to improve the detection accuracy of the model.To address the challenges posed by the diverse features of biscuit surfaces,including unclear or insufficient surface defect features and the presence of edible particles,this thesis introduces the sc SE attention mechanism to focus on important feature regions of targets,thus enhancing feature extraction capabilities and further improving detection accuracy.In addition,to solve the problem of missed detection of small targets such as hair,this thesis uses the structure of Bi FPN to fuse features more effectively and retain more detailed information,thus further improving the accuracy of the detection model.Finally,the YOLOv5-KSB surface defect detection model is formed.(4)Based on the YOLOv5-KSB framework,a biscuit surface defect detection platform has been designed,which can automatically detect biscuit surface defects and display the detection results.The platform supports a variety of input modes,including pictures and videos in file input mode,and direct connection of cameras to meet the production needs of different users.In addition,it also provides a threshold and confidence setting interface to meet the requirements of different users for recognition accuracy.