4 results about "Network complexity" patented technology

The application provides a feeding network, an antenna module and an apparatus. In the feeding network, the feeding network comprises M radio frequency interfaces, N antenna interfaces, at least M first distribution units, N second distribution units and Mx(N-1) phase shift units, M is less than or equal to N, and M and N are integers greater than 1; each radio frequency interface is configured to receive a first radio frequency signal; the first distribution unit is configured to process the first radio frequency signal into N second radio frequency signals; an input end of each second distribution unit is connected with at least one first distribution unit and / or at least one phase shift unit, and an output end of the second distribution unit is connected with the antenna interface; each phase shift unit is connected between the first distribution unit and the second distribution unit, and each phase shift unit is configured to adjust the phase of the received second radio frequency signal. In this way, the network complexity of the feeding network is low, the insertion loss is small, and the isolation is good.

The application discloses a kind of underwater target detection method and device based on dynamic cascade refinement and multi-granularity feature aggregation, and relates to underwater target detection technical field.The method comprises: by introducing dynamic cascade feature refinement module and focal modulation module, effectively enhance the capture ability of network to fuzzy and small target detail features.Context attention aggregation and efficient feature enhancement module are embedded in neck network, information fusion across channels and across spaces is realized, and the richness and accuracy of feature expression are significantly improved.Combined with the design of layered adaptive unified detection head, the model reduces network complexity while ensuring computational efficiency.The method has the advantages of rapid, lightweight and high accuracy, can effectively eliminate redundant frame and accurately output target position, and is suitable for real-time detection application in various underwater scenes.

This invention discloses a method for performance evaluation and optimization of automatic differential operators. Under specified hardware and a deep learning framework, it evaluates multiple candidate automatic differential operators and constructs performance profiles. Combining the target scenario, it sets data batch anchor points, compares and analyzes to determine the optimal operator at each anchor point, and uses this operator to train a physical information neural network, gradually increasing the network complexity. It then obtains the maximum runnable network structure under hardware constraints and the upper bound of the data batch without memory overflow, generating corresponding recommended configurations. This invention achieves reproducible operator performance profiles under unified evaluation, quantitatively assesses computational efficiency and resource consumption under different data and network scales, and provides a quantifiable optimization basis for the training configuration of physical information neural networks.

The application provides an implementation method of a raw domain knowledge distillation lightweight denoising network, which comprises the following steps: firstly, selecting a suitable convolution layer output feature as a guide layer from a teacher network layer which has been trained and frozen through a feature distillation mode; then selecting a convolution output of a corresponding layer from a student model which has been constructed; finally, calculating a total loss of a loss of the corresponding layer and a loss between an output of the student model itself and a clean image, and using the total loss to act on the student model, so that the student model is iterated to a stable result according to the loss. The method uses a compression strategy of feature distillation, can transfer the feature information extraction capability of a large model to a student model, and ensures that the denoising effect of the small model is not affected while the network complexity is reduced.