The identification of handwritten characters has been a major focus of research, owing to the ease of use and improved language comprehension provided by technological advancements. In this study, we present a model for detecting compound Bangla characters from handwritten images. Of the almost 320 compound characters in the Bangla language, 171 have been previously studied by other researchers. We focused on the remaining 50 characters that have not been studied before. Our model's accuracy is dependent on the quality and comprehensiveness of the dataset, which we collected by gathering data on sheets and converting it to JPG format. To process the images, we developed a box detection algorithm in Python open CV, along with other processing techniques. Our machine learning model used a multilayer CNN architecture comprising convolution, maxpooling, dense, and dropout layers. The model achieved an accuracy of 88.48%, which is considered good. This study contributes to the field of handwritten identification for the Bangla language and demonstrates the efficacy of our approach in detecting previously unstudied characters.