Comparison of convolutional neural network models for food image classification

Figures & data

Figure 1. An example of convolution step.

Figure 2. AlexNet Architecture (Krizhevsky, 2012; Karnowski, 2015).

Figure 3. Samples from the dataset.

Table 1. Proposed structures.

Structure 1	Structure 2	Structure 3
Conv1: 7: 4: 0: 120: ReLU Max Pool:3:2 LRN	Conv1: 11: 4: 0: 96: ReLU Max Pool:3:2	Conv1: 7: 4: 0: 120: ReLU Max Pool:3:2
Conv2: 5: 1: 2: 240: ReLU Max Pool:3:2 LRN	Conv2: 5: 1: 2: 256: ReLU Max Pool: 3:2	Conv2: 5: 1: 2: 240: ReLU Max Pool:3:2
Conv3: 3: 1: 1: 360: ReLU	Conv3: 3: 1: 1: 384: ReLU	Conv3: 3: 1: 1: 360: ReLU
Conv4: 3: 1: 1: 360: ReLU	Conv4: 3: 1: 1: 384: ReLU Max Pool:3:2	Conv4: 3: 1: 1: 360: ReLU
Conv5: 3: 1: 1: 360: ReLU Max Pool:3:2	FC1: 4096: ReLU Dropout	Conv5: 3: 1: 1: 360: ReLU Max Pool:3:2
FC1: 500: ReLU Dropout	FC2: 4096: ReLU Dropout	FC1: 500: ReLU Dropout
FC2: 300: ReLU Dropout	FC3: Softmax	FC2: 300: ReLU Dropout
FC3: Softmax		FC3: Softmax

Table 2. Learning parameters.

Parameters	Values
Learning rate	0.0001
Momentum	0.9
Maximum iteration	600 (Adam), 2000 (SGD), 600 (Nesterov’s)

Table 3. Test results.

Model	Accuracy (%) (SGD)	Accuracy (%) (Nesterov’s accelerated gradient)	Accuracy (%) (Adam)
CaffeNet (BVLC-Caffenet)	80.51	81.55	83.7
AlexNet (BVLC-Alexnet)	82.07	83.01	86.92
Structure 1	70.12	71.55	73.80
Structure 2	38.85	47.63	71.7
Structure 3	53.78	68.55	69.92

Karnowski, J. (2015). AlexNet + SVM. Retrieved from https://jeremykarnowski.files.wordpress.com/2015/07/alexnet2.png

 Google Scholar