We propose a deep learning-based sensor that mitigates the problem of crystal detection in high-density slurries using a segmentation model based on the RetinaNet framework. The sensor functions in a single stage in contrast to current state-of-the-art deep learning models such as the Mask R-CNN that require two stages. It does this by dividing the work among three subnetworks that work in parallel with each one solving a slightly different problem. While the first subnetwork localizes objects in the image, the second predicts their class labels, with the final task of generating pixel-maps handled by the third. The three pieces of information are then combined and processed to quantify the shape characteristics of the crystals and to generate their size distribution. Because of a reduction in the number of stages, the sensor has fewer parameters than the Mask R-CNN (36 M vs 44 M) without compromising on the quality of the results. The performance of the sensor is evaluated on an experimental study involving the antisolvent crystallization of the NaCl-ethanol-water system.