The problem of clustering data arrays described in both vector and matrix forms and based on the optimization of data distribution density functions in these arrays is considered. For the optimization of these functions, the algorithm that is a hybrid of Fish School Search, random search, and evolutionary optimization is proposed. This algorithm does not require calculating the optimized function’s derivatives and, in the general case, is designed to find optimums of multiextremal functions of the matrix argument (images). The proposed approach reduces the number of runs of the optimization procedure, finds extrema of complex functions with many extrema, and is simple in numerical implementation.