Complex Labels Text Detection Algorithm Based on Improved YOLOv5.

Complex labels have been widely used in various industries. The accuracy of its content is critical both in the fields of people's livelihood, such as supermarkets and shopping centers, and in the management of goods in the industrial fields, such as logistics and factories. Inaccurate label information identification can make item management difficult. Because complex labels can simultaneously contain text, icons, bar codes, QR codes, and other information with different aspect ratios. Traditional methods like feature extraction and template matching have problems, such as detection frames breaking between Chinese, English, and numeric symbols. As a result, entire lines of text on complex labels cannot be detectable, resulting in low detection accuracy. In this paper, a deep learning-based text detection algorithm was proposed. By employing the operation of inverse convolution, improved the object detection algorithm you only look once 5 (YOLOv5). In the backbone part of the original model, involution is used instead of the convolution layer to improve target classification and prediction. The original anchor frame was modified using k-means clustering to make it more applicable to text of various sizes in labels. The enhanced algorithm is called as Invo-YOLOv5. Experiments show that this model can significantly improve detection efficiency while also addresse the problems of false detection and missed detection. Finally, the detected text is verified by using CRNN and Tesseract OCR with complex labels as samples for recognition. Both methods can be effectively recognized, demonstrating the efficacy and generality of the Invo-YOLOv5 method in the process of complex labels text detection and improving the detection accuracy. [ABSTRACT FROM AUTHOR]