Several fields most notably object identification, have effectively employed artificial intelligence (AI). Nevertheless, deep learning (DL) models are costly in terms of memory and computational power, making it difficult to apply them in systems that must react quickly to disruptions or have restricted resources. Consequently, in implementing these deep models on edge devices lacking appreciably compromising their performance, they must be expedited and compacted to lower dimensions. In this manuscript, novel compression approaches of the DL technique are investigated and compared based on the pruning, quantization, and knowledge distillation (KD) models, an efficaciously developing method in this field. Initially, the images collected from the raw database are denoised using the Extended Kuan filtering (Ex-KF) technique to achieve better detection performance. Then, the convoluted-duo-directed long short-term memory (Conv-DDLSTM) model is introduced, and various compression processes are applied to minimize model complexities. Moreover, the proposed framework is deployed on edge mobile devices for automatic object detection (OD). The experimentation is tested with the Python Platform and a freely accessible COCO-2017 database is utilized. Various assessment measures like accuracy, F-measure, Matthew’s Correlation Coefficient (MCC), positive predictive value (PPV), Intersection of Union (IoU), mean average precision (mAP), power and memory consumption, model size, and latency are scrutinized with other conventional frameworks. The proposed method achieved overall accuracy of 98.68% and 97.93%, and power consumption of 216.70 MW and 163.25 MW for the presence and absence of compression respectively.
More et al. (Fri,) studied this question.