Pharmacists in conventional pharmacies and hospitals prepare and visually inspect medications before dispensing these medications to patients, thereby posing risks of exposure through skin contact or inhalation, as well as the potential for generating compounding errors such as incorrect medication types or quantities. To address the limitations of the manual compounding processes, automated compounding and medication-verification systems have been introduced and implemented. The automated compounding system prepares medications in pouch units based on prescriptions written by doctors, whereas the medication verification system verifies whether the types and quantities of the medication dispensed in these pouches match the prescription information. Medications are categorized broadly into solid tablets, hard capsules, and soft capsules. Color and shape information is used to classify medication types; solid tablets and hard capsules can be classified using rule-based image processing techniques such as image binarization. However, extracting shape information from soft capsules through which light passes is impossible via image binarization alone; therefore pattern matching is employed. Nevertheless, soft capsules change their forms, causing significant variations in pattern-matching similarity; therefore the detection accuracies of these capsules are low. To address this issue, this paper proposes a deep learning-based method for detecting soft capsules.