When the intelligent inspection machine extracts the impurity in infusion liquid, because of the interference of image displacement deviation, the misjudgment phenomenon always occurs when using the frame difference method to detect the impurity. In order to solve the problem, a new method of binary descriptor block matching was proposed based on Features of Accelerated Segment Test (FAST). Firstly, the feature points were detected by accelerating the segment test on different scales of the image, and the best feature point was chosen by using non-maximal suppression and the entropy difference. Then, the improved template was used for sampling around the feature point, which formed the new binary descriptor with strong robustness to scale changes, noise interference and illumination changes. The dimension of new descriptor was further reduced. Finally, by using the block matching and threshold method, the two frame images were matched quickly and accurately, and the displacement deviation was solved and compensated. The experimental results show that, when processing the 1.92 million pixel image, the overall real-time performance of the proposed method can be up to 190 ms, and the new descriptor generation only accounts 96 ms. The matching accuracy of the proposed algorithm is more than 99%, which suppresses the error matching of large spatial position offset successfully. The calculated deviation error of the proposed method is much less than the existing algorithms of Scale Invariant Feature Transform (SIFT) and ORiented Binary robust independent elementary features (ORB) with high matching precision. And with the displacement compensation which can be accurate to sub-pixel level, the proposed method can rapidly compensate the displacement deviation of the bottle in the image.

For the problem of existing models' disability to analyze topic-sentiment evolution of microblogs, a Dynamic Topic Sentiment Combining Model (DTSCM) was proposed based on Topic Sentiment Combining Model (TSCM) and the emotional cycle theory. DTSCM could track the topic sentiment evolution trend and obtain the graph of topic sentiment evolution so as to analyze the evolution of topic and sentiment by capturing the topic and sentiment of microblogs in different time. The experimental results in real microblog corpus showed that, in contrast with state-of-the-art models Joint Sentiment/Topic (JST), Sentiment-Latent Dirichlet Allocation (S-LDA) and Dependency Phrases-Latent Dirichlet Allocation (DPLDA), the sentiment classification accuracy of DTSCM increased by 3.01%, 4.33% and 8.75% respectively,and DTSCM could obtain topic-sentiment evolution of microblogs. The proposed approach can not only achieve higher sentiment classification accuracy but also analyze topic-sentiment evolution of microblog, and it is helpful for public opinion analysis.