It is widely observed in the industry that defective dies tend to occur in groups of systematic pattern. These are so-called defect clusters. There are many proposed methods to achieve cluster classification and recognition with different degree of accuracy and limitations. Many of these methods, although powerful, generally do not actually detect the presence/absence of a cluster but simply segments them and then attempts to calculate the validity of the segment. Thus, they fail to be flexible and accurate because they implicitly assume that the problem is singular: identify the defect clusters, when in actuality, the problem of defect cluster identification can be divided into three distinct stages: detection, segmentation and recognition. This paper proposes the use of joint-count statistics to perform the sole task of defect cluster detection. It is recommended that segmentation and recognition be performed after the detection algorithm completed to a satisfactory level.