Research On Fabric Defect Small Object Detection Technology Based On Deep Learning

For textile enterprises,improving production efficiency and reducing production costs are important goals,but ensuring the quality of fabric products is a prerequisite for achieving this goal.Although modern and advanced textile machinery can effectively reduce the appearance of fabric defects,the various defects caused by the manufacturing process are still unavoidable.At present,the inspection tasks of fabric defects are mostly done manually,and this labor-intensive work does not guarantee the consistency evaluation of products.Aiming at the problems in fabric defect detection method,this paper proposes a fabric-based small object detection technology based on machine learning and computer vision to realize real-time detection and accurate positioning of fabric defects,which can effectively improve the accuracy and efficiency of fabric defect detection.The main work of this paper includes the following points:(1)Build an image acquisition platform,collect fabric defect images in real time,and construct a fabric defect data set for the pre-processed images.(2)In view of the shortcomings of SSD model and DSSD model in object detection application,the two networks are improved,and a Multi-scale Fusion Deconvolutional Single Shot Detector(MFDSSD)fabric small object detection method based on deep learning is proposed.The multi-layer feature extraction and fusion strategy is integrated into the MFDSSD network,and the a priori frame of the SSD network is optimized to improve the detection capability of the model for small size defects.Then train and test the network to achieve fast and stable fabric picking and detection.(3)Design comparison experiments to analyze the detection performance of the MFDSSD model in different fabric types.The evaluation system of the model is established and compared with the common object detection algorithms in different fabric data sets.The evaluation results are obtained and the reliability of the algorithm is verified.The experimental results show that the average detection accuracy of the MFDSSD algorithm for fabric solid color cloth and printed fabric is 82.8% and 74.1%,respectively,and the detection speed is 52 FPS and 46.2FPS respectively.Experiments prove that the MFDSSD algorithm achieves a balance between detection time and positioning accuracy.(4)Design fabric inspection software system,including online detection and offline detection.The system can realize picture acquisition,image preprocessing,defect detection,and can save test results.The MFDSSD algorithm proposed in this paper can detect fabric defects of various shapes and backgrounds,effectively reducing the number of missed detecti ons and false reductions.The use of software systems can help fabric manufacturers improve product quality.