| With the development of object detection algorithms in deep learning,a new automated inspection method has emerged in industrial product quality control.This method has improved the efficiency and accuracy of industrial product quality inspection and has gradually become the primary means of industrial quality control.However,existing object detection algorithm research mainly focuses on conventional scenes and lacks research on the industrial product with complex textures.Compared to regular industrial products,complex textures on industrial products have two significant characteristics.Firstly,detecting small-sized defects is challenging,and there are a large number of low-visibility small-sized defects in detection scenarios on complex textured industrial surfaces often involve.Secondly,complex textured industrial products are heavily affected by background interference,causing current detection algorithms to mistakenly identify normal texture variations as product defect areas.Therefore,to enhance the efficiency of object detection algorithms in detecting surface defects on complex textured industrial products,the details are divided into the following parts:(1)Complex texture surface defect detection method based on channel and spatial attention.To improve the detection accuracy of surface defects in complex textured industrial products,this thesis focuses on the challenges of detecting small-sized defects that are difficult to detect and the significant background interference in complex textured industrial products.Firstly,a selective feature fusion method is proposed to overcome the problem of low visibility and difficulty in detecting small-sized defects.In the process of deep and shallow feature fusion,an additional branch is introduced to dynamically generate feature fusion weights by modeling the relationship between deep and shallow feature channels,thus achieving selective feature fusion.Secondly,to suppress background interference from complex textures,a channel and spatial joint attention method is designed.Based on the morphology and distribution characteristics of surface defects in textured industrial products surface,channel attention is used to model the local relationship of feature channels and improve the weights of key feature channels in model training.Spatial attention is applied to capture long-range relationships and selectively aggregate features in key regions.Finally,the effectiveness of this approach is verified through ablation experiments on a dataset of textured ceramic tile surface defects.(2)Automated detection system for complex texture ceramic tile surface defects.To further reduce the difficulty of detecting complex textured industrial products in actual quality control processes,a detection system for surface defects on complex texture ceramic tiles is designed and implemented based on the aforementioned method.In addition to the conventional features of defect visualization,data annotation,and data management in a typical detection system,this thesis considers the quality of complex textured industrial product images and incorporates an image acquisition and image preprocessing module into the system.This module aims to provide highdefinition images of complex textured industrial products and fully utilize the highresolution image details,thereby reducing the difficulty of the actual quality control process. |
