Abstract: In industrial defect detection, the massive computational overhead of deep neural networks (DNNs) makes it difficult to satisfy the hard real-time constraints of edge devices. Furthermore, existing methods are typically limited to single-dimensional scaling and lack joint optimization for heterogeneous multi-core architectures. To address this issue, we propose a joint optimization and scheduling framework for dynamic DNNs oriented toward heterogeneous multiprocessor systems. At the network layer, we design a dynamic DNN architecture that supports the three-dimensional collaborative scaling of depth, width, and resolution, alongside an incremental training strategy. This significantly enhances both the model's inference accuracy and training efficiency. At the scheduling layer, a priority dual-queue mechanism based on image block weighting is introduced. The heterogeneous task scheduling is formulated as a multi-level generalized assignment problem, and a heuristic algorithm based on linear relaxation and bidirectional dynamic adjustment (LRBDA) is proposed to solve the heterogeneous scheduling problem. Experimental results demonstrate that the proposed hybrid-scaling architecture significantly outperforms single-dimensional scaling strategies and existing multi-branch dynamic networks in balancing accuracy and computational cost, reducing the computational overhead by up to 23.54% under the same detection accuracy. Furthermore, compared to the mixed-criticality EKG scheduling algorithm (MC-EKG) algorithm, the LRBDA algorithm overcomes the performance bottleneck associated with long periods, significantly improving the average system accuracy and schedulability range under stringent computational constraints.