To identify and solve this problem, you could use OCR and anomaly detection. Firstly, fetch structured data from the images using Tesseract OCR or google Vision API then, clean and organize them in data frame. The usage of statistical methods like mean, IQR and standard deviation or ML models like Isolation, Forest, Autoencoders, Clustering can detect unusual number of event counts. Categorize zero values by comparing them with typical patterns of the past time period to determine if they are intended or error, At the very end, weed out wrong-flagged anomalies and enhance the detection model.