Combining MobileNetV1 and Depthwise Separable convolution bottleneck with Expansion for classifying the freshness of fish eyes

• Published in: Information processing in agriculture 2022-12, Vol.9 (4), p.485-496
• Main Authors: Prasetyo, Eko, Purbaningtyas, Rani, Adityo, Raden Dimas, Suciati, Nanik, Fatichah, Chastine
• Format: Article
• Language: English
• Subjects: Bottleneck; Classification; Depthwise separable convolution; Fish eye; Freshness; Residual transition
• ISSN: 2214-3173; 2214-3173
• DOI: 10.1016/j.inpa.2022.01.002

•There are not enough appearance features in classifying the freshness of fish eyes.•Slight differences of images do not require a complicated CNN architecture.•We proposed Depthwise Separable Convolution Bottleneck with Expansion (DSC-BE) to generate features for differentiating fresh and not fresh.•We proposed Residual Transition (RT) to bridge current feature maps and skip connection features.•We proposed MobileNetV1 with Bottleneck and Expansion (MB-BE) consists of MobileNetV1, DSC-BE, and RT. Image classification using Convolutional Neural Network (CNN) achieves optimal performance with a particular strategy. MobileNet reduces the parameter number for learning features by switching from the standard convolution paradigm to the depthwise separable convolution (DSC) paradigm. However, there are not enough features to learn for identifying the freshness of fish eyes. Furthermore, minor variances in features should not require complicated CNN architecture. In this paper, our first contribution proposed DSC Bottleneck with Expansion for learning features of the freshness of fish eyes with a Bottleneck Multiplier. The second contribution proposed Residual Transition to bridge current feature maps and skip connection feature maps to the next convolution block. The third contribution proposed MobileNetV1 Bottleneck with Expansion (MB-BE) for classifying the freshness of fish eyes. The result obtained from the Freshness of the Fish Eyes dataset shows that MB-BE outperformed other models such as original MobileNet, VGG16, Densenet, Nasnet Mobile with 63.21% accuracy.