Concerning the poor performance of the Single-Shot Refinement Neural Network for Object Detection (RefineDet) of the object detection network when detecting small sample classes in inter-class imbalanced datasets, a Subsection Weighted Loss (SWLoss) function was proposed. Firstly, the inverse of the number of samples from different classes in each training batch was used as the heuristic inter-class sample balance factor to weight the different classes in the classification loss, thus strengthening the concern on the small sample class learning. After that, a multi-task balancing factor was introduced to weight classification loss and regression loss to reduce the difference between the learning rates of two tasks. At last, experiments were conducted on Pascal VOC2007 dataset and dot-matrix character dataset with large differences in the number of target class samples. The results demonstrate that compared to the original RefineDet, the SWLoss-based RefineDet clearly improves the detection precision of small sample classes, and has the mean Average Precision (mAP) on the two datasets increased by 1.01 and 9.86 percentage points, respectively; and compared to the RefineDet based on loss balance function and weighted pairwise loss, the SWLoss-based RefineDet has the mAP on the two datasets increased by 0.68, 4.73 and 0.49, 1.48 percentage points, respectively.

Aiming at the ignorance of target structure in test topic distribution due to the lack of annotation in supervised Latent Dirichlet Allocation (sLDA) model, a sLDA fastener image classification model combined with global and local constraints (glc-LDA) was proposed. Firstly, the training images were manually labeled, and the training topic distribution with structural information was learned in sLDA model. Then, the test topic distribution was calculated to obtain the image category probabilities as global constraints, the topic similarities of training sub-blocks and test sub-blocks as local constraints. Finally, updated test topic distribution was obtained by weighted summation of training topic distribution with the product of global and local constraints as updated weights. The image category labels were obtained in Softmax classifier by the updated topics. The experimental results show that the proposed algorithm can express the structural information of fastener and compared with sLDA model, the distinction of each category of fastener images is enhanced, and the false detection rate is reduced by 55%.

Focusing on the ignorance of the image structure in Latent Dirichlet Allocation (LDA) model, a LDA fastener detection model integrated with image texture information, namely TS_LDA, was proposed. Firstly, a single-channel Local Binary Pattern (LBP) method was designed to acquire the image texture structure, and the texture information was treated as a label of visual word. The joint distribution of words and labels reflected the characteristics of an image structure. Secondly, those labels were embedded into LDA, and image topics were derived from words and labels. The improved distribution of topics considered the image structure. Finally, the classifier was trained and fastener states were identified on the basis of topic distribution. Compared with the LDA method, the differences between normal and disabled fasteners increased by 5%-35%, the average misdetection rate decreased by 1.8%-2.4% in the topic space of TS_LDA. The experimental results show that TL_LDA is able to enhance the accuracy of fastener image modeling, thus inspects fastener states more precisely.