Detection of cigarette appearance defects based on improved SSD model

Detecting the tiny defects of cigarette appearance based on a hybrid model of lightweight ViT and RCNN

“…The Adam optimization technique was used to replace the momentum approach, greatly reducing the computational effort and amount of parameters required to detect cigarette capsules quickly and in real-time. The base network of SDD, VGG16, is replaced with a ResNet50 network, which has improved characterization capabilities, by Qu et al [5]. Utilizing feature pyramid convolution and loss function optimization, faults in the appearance of cigarettes were detected.…”

“…In conclusion, CNN and Transformer object detection has been widely used in manufacturing operations. Similar to the work of Li [11] and Qu [5], both of which address domain problems in certain production processes, this paper applies the Swin Transformer model for flaw detection in the appearance of cigarettes. The differences include 1) the scale of the solution, the object scale studied in this paper is in the range of 1%, and 2) the high real-time requirements, applying Swin Transformer and Mask RCNN hybrid lightweight models to meet the real-time requirements.…”

“…For example, Xiao [3] and Yang [4] used a segmented area ratio method for cigarette defects detection, which has the disadvantage of not being effective in detecting tiny cigarette objects. The speed and accuracy of deep learning-based CNN models have greatly improved, but high accuracy requires large-scale data for training and a reasonable supervised algorithm, For example, Qu et al [5] improved SSD networks using machine learning and CNN methods to detect cigarette defects, but the model is difficult to meet the actual production requirements for tiny object detection in terms of comprehensive performance and accuracy. The application of Vision Transformer (ViT) in computer vision [6] provides a novel method for detecting flaws in cigarettes.…”

An Appearance Defect Detection Method for Cigarettes Based on C-CenterNet

“…It can be concluded from the analysis of the results in Table 2 that, although the method proposed in this paper is not optimal in terms of average detection speed and is slower than YOLOv5, the accuracy rate, recall rate, and mAP are the highest in the table, indicating that the method yields fewer misses and is more accurate. Although the detection speed is not as good as that of YOLOv5, the detection speed achieved by the method in this paper can meet the detection needs of most current situations [35], and is much faster than the detection speed of existing research [22,23]. In addition, compared with CenterNet, the method demonstrates 7.24% higher accuracy, 7.66% higher recall, and 6.14% higher mAP, and the average detection time was increased by only 1.2 ms/branch.…”

“…Although neural networks have been proven to be effective in many detection tasks, there are still many challenges related to the detection of cigarette appearance defects: the data set of cigarette appearance defects is insufficient, the defect scale changes greatly, and the cigarette image is narrow and long. Qu et al improved the SSD network with methods such as pyramid convolution to realize the detection of cigarette appearance defects with complex features and imbalanced datasets [22]. Li et al used MobileNet to replace Vgg16 of traditional Faster R-CNN to effectively detect cigarette capsule defects [23].…”