Table 1 Comparison of different algorithms for KFE.
From: Key frame extraction algorithm for surveillance videos using an evolutionary approach
Algorithm | Architecture/methodology | Limitations |
|---|---|---|
DL | DL such as Capsule Net, YOLO, LSTM, Deep Summarization Network (DSN), Fully Connected Sequence Network, Attention-based Encoder, Decoder Network, RNN, Chunk and Stride Net (CSNet), Summary Net, Detect to Summarize Net (DSNet) were investigated | These models had been evaluated exclusively on specific datasets They are computationally expensive as they require powerful GPUs and a significant amount of training time The DL requires fine-tuning many hyperparameters, such as learning rate, model size, and regularization techniques. This process can be challenging and frequently involves much trial and error, unlike GA’s, which are self-adaptive |
Image Processing and Machine Learning (ML) | K-Means Clustering With DCT, Sparse Dictionary Selection, Semantic Graphs, Odometry Estimation By Scan Line Similarity, SIFT, Color SIFT, HWVP Descriptors, GIST, HSV, PHOG Descriptors With K-Means, SSPA Curve, LBP, Optical Flow | They are designed for specific applications Model and hyperparameters limit the ML. Image processing is task-specific and non-adaptable. Unlike GA’s that are highly adaptive and evolve dynamically |
Evolutionary Approaches | Application of evolutionary algorithms for Image segmentation7, Image fusion8, and for solving COP’s9,14 GA-based KFE | The FF involved in GA for KFE is just based on color distance It lacks generality and does not apply to all types of videos There is no effective survivor selection mechanism |
