基于改进YOLOv7的织物疵点检测算法

Fabric defect detection algorithm based on improved YOLOv7

为解决因织物疵点尺寸小、形状不规则而导致检测准确率较低等问题,提出一种基于YOLOv7的改进算法。首先设计一个新的聚合网络DR-SPD,该结构将DRes的动态感知能力和SPD的细节提取能力结合在一起,在保持动态区域视野的同时减少细粒度信息的丢失;针对织物背景对检测效果造成的影响,引入GAM注意力机制,提升模型的抗干扰能力,使其能聚焦于更为关键的语义信息;最后,在特征融合网络中加入3条横向跳跃路径以缩短深浅层之间信息传递的距离,减少细节特征的遗漏率。试验表明:改进后模型的mAP值达到95.63%,检测速度为51帧/s,综合性能优于其他主流模型。为验证在工业场景中的有效性,将改进模型部署到车间设备上进行测试,其mAP值达到94.85%,检测速度为43帧/s,可满足实际生产需求。

To solve the problem of low detection accuracy caused by small size of fabric defects and irregular shape,an improved algorithm based on YOLOv7 was proposed.Firstly,a new aggregation network DR-SPD was designed,which combined the dynamic perception ability of DRes with the detail extraction ability of SPD to reduce the loss of fine-grained information while maintaining a dynamic regional view.Aiming at the influence of fabric background on detection effect,the GAM attention mechanism was introduced to enhance the model's antiinterference ability,enabling it to focus on more critical semantic information.Finally,three lateral jump paths were added to the feature fusion network to shorten the distance of information transmission between deep and light layers,reduce detail omissions of detail feature.Experiments showed that mAP value of the improved model was reached 95.63%,the detection speed was 51 flame/s and the overall performance was better than other mainstream models.To verify its effectiveness in industrial scenarios,the improved model was deployed on workshop equipment for testing.The mAP value was reached 94.85%and the detection speed was reached 43 flame/s.It could meet the actual production needs.