Droplet-based dPCR offers many advantages over chip-based dPCR, such as lower processing cost, higher droplet density, higher throughput, while requiring less sample. However, the stochastic nature of droplet locations, uneven illuminations, and unclear droplet boundaries make automatic image analysis challenging. Most methods currently used to count a large amount of microdroplets rely on flow detection. Conventional machine vision algorithms cannot extract all information of the targets from complex backgrounds. Some two-stage methods, which first locate and then classify droplets according to their grayscale values, require high-quality imaging. In this study, we addressed these limitations by improving a one-stage deep learning algorithm named YOLOv5 and applying it to the detection task to realize one-stage detection. We introduced an attention mechanism module to increase the detection rate of small targets and used a new loss function to speed up the training process. Furthermore, we employed a network pruning method to facilitate the deployment of the model on mobile devices while preserving its performance. We validated the model with captured droplet-based dPCR images and found that the improved model accurately identified negative and positive droplets in complex backgrounds with an error rate of 0.65%. This method is characterized by its fast detection speed, high accuracy, and ability to be used on mobile devices or cloud platforms. Overall, the study presents a novel approach for detecting droplets in large-scale microdroplet images and provides a promising solution for accurate and efficient droplet counting in droplet-based dPCR.